@jax-js/jax 0.1.8 → 0.1.10

package/dist/{webgpu-BykvF26B.cjs → webgpu-DMSx7a6M.cjs}

@@ -1,4 +1,4 @@
-const require_backend = require('./backend-B3foXiV_.cjs');
+const require_backend = require('./backend-DMauYnfl.cjs');
 
 //#region src/backend/webgpu/builtins.ts
 const threefrySrc = `
@@ -152,6 +152,7 @@ var SyncReader = class SyncReader {
 		this.device = device;
 	}
 	#init() {
+		if (typeof OffscreenCanvas === "undefined") throw new Error("OffscreenCanvas is not available in this environment, so you cannot read data from WebGPU synchronously. Consider using the async API.");
 		const makeCanvas = () => new OffscreenCanvas(SyncReader.width, SyncReader.height);
 		this.deviceStorage = SyncReader.alphaModes.map(makeCanvas);
 		this.deviceContexts = this.deviceStorage.map((canvas, i) => {
@@ -247,6 +248,10 @@ function bitonicSortUniform(pass) {
 *   `2^(step+1)` with multiple workgroups. This doesn't use shared memory.
 *
 * The total number of passes is roughly `log2(n / workgroupSize)^2 / 2`.
+*
+* If `outputIndices` is true, the shader also tracks the original indices of
+* the sorted elements (argsort) and outputs them to a separate buffer. This
+* also makes the sorting algorithm stable.
 */
 function bitonicSortShader(device, dtype, n, batches, outputIndices) {
 	const ty = dtypeToWgsl(dtype, true);
@@ -286,14 +291,21 @@ ${require_backend.isFloatDtype(dtype) ? `
 fn compare_and_swap(i: u32, j: u32) {
   let val_i = shared_vals[i];
   let val_j = shared_vals[j];
-  if (compare(val_j, val_i)) {
+${outputIndices ? `
+  if (
+    compare(val_j, val_i) ||
+    (!compare(val_i, val_j) && shared_idx[j] < shared_idx[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;` : ""}
-  }
+    shared_idx[j] = tmp_idx;
+  }` : `
+  if (compare(val_j, val_i)) {
+    shared_vals[i] = val_j;
+    shared_vals[j] = val_i;
+  }`}
 }
 
 @compute @workgroup_size(${workgroupSize})
@@ -370,13 +382,17 @@ ${outputIndices ? `
     if (j < ${n}u) {
       let val_i = output[base + i];
       let val_j = output[base + j];
-      if (compare(val_j, val_i)) {
+${outputIndices ? `
+      let idx_i = output_idx[base + i];
+      let idx_j = output_idx[base + j];
+      if (compare(val_j, val_i) || (!compare(val_i, val_j) && idx_j < idx_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;` : ""}
+        output_idx[base + i] = idx_j;
+        output_idx[base + j] = idx_i;` : `
+      if (compare(val_j, val_i)) {
+        output[base + i] = val_j;
+        output[base + j] = val_i;`}
       }
     }
   }
@@ -713,6 +729,120 @@ function createRoutineShader(device, routine) {
 	}
 }
 
+//#endregion
+//#region src/backend/webgpu/tracing.ts
+const MAX_TIMESTAMP_QUERIES = 4096;
+const activeBatch = /* @__PURE__ */ new WeakMap();
+function createTracingBatch(device) {
+	return {
+		querySet: device.createQuerySet({
+			type: "timestamp",
+			count: MAX_TIMESTAMP_QUERIES
+		}),
+		resolve: device.createBuffer({
+			size: MAX_TIMESTAMP_QUERIES * 8,
+			usage: GPUBufferUsage.QUERY_RESOLVE | GPUBufferUsage.COPY_SRC
+		}),
+		dst: device.createBuffer({
+			size: MAX_TIMESTAMP_QUERIES * 8,
+			usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST
+		}),
+		nextIndex: 0,
+		entries: []
+	};
+}
+function acquireTracingSlot(device) {
+	if (!device.features.has("timestamp-query")) return void 0;
+	let batch = activeBatch.get(device);
+	if (batch && batch.nextIndex >= MAX_TIMESTAMP_QUERIES) {
+		flushTracingBatch(device, batch);
+		batch = void 0;
+	}
+	if (!batch) {
+		batch = createTracingBatch(device);
+		activeBatch.set(device, batch);
+		require_backend.onFlushTrace(() => {
+			const b = activeBatch.get(device);
+			if (b && b.entries.length > 0) flushTracingBatch(device, b);
+			activeBatch.delete(device);
+		});
+	}
+	const beginIndex = batch.nextIndex;
+	const endIndex = beginIndex + 1;
+	batch.nextIndex += 2;
+	return {
+		batch,
+		beginIndex,
+		endIndex
+	};
+}
+/**
+* If tracing is active, acquire a slot for timestamp queries.
+*
+* Returns undefined if tracing is not active or the device doesn't support
+* timestamp queries.
+*/
+function maybeAcquireTracingSlot(device) {
+	if (!require_backend.isTracing()) return void 0;
+	return acquireTracingSlot(device);
+}
+/**
+* Record a tracing entry for a pipeline dispatch and schedule an auto-flush.
+*/
+function recordTrace(device, slot, source, numPasses, wgslSource) {
+	const info = require_backend.traceSourceInfo(source);
+	info.properties.push(["passes", `${numPasses}`]);
+	info.properties.push(["source", wgslSource]);
+	slot.batch.entries.push({
+		...info,
+		beginIndex: slot.beginIndex,
+		endIndex: slot.endIndex
+	});
+	scheduleAutoFlush(device);
+}
+/**
+* If the active batch has pending entries, flush and replace it so traces
+* are emitted without waiting for the batch to fill or stopTrace().
+*
+* Called after each dispatch records its entry via a microtask so that
+* synchronous back-to-back dispatches are still batched together.
+*/
+function scheduleAutoFlush(device) {
+	queueMicrotask(() => {
+		const batch = activeBatch.get(device);
+		if (batch && batch.entries.length > 0) {
+			flushTracingBatch(device, batch);
+			activeBatch.set(device, createTracingBatch(device));
+		}
+	});
+}
+function flushTracingBatch(device, batch) {
+	if (batch.entries.length === 0) return;
+	const usedQueries = batch.nextIndex;
+	const encoder = device.createCommandEncoder();
+	encoder.resolveQuerySet(batch.querySet, 0, usedQueries, batch.resolve, 0);
+	encoder.copyBufferToBuffer(batch.resolve, 0, batch.dst, 0, usedQueries * 8);
+	device.queue.submit([encoder.finish()]);
+	const { entries } = batch;
+	batch.dst.mapAsync(GPUMapMode.READ).then(() => {
+		try {
+			const times = new BigInt64Array(batch.dst.getMappedRange());
+			const anchorGpuNs = times[entries[entries.length - 1].endIndex];
+			const anchorCpuMs = performance.now();
+			for (const entry of entries) {
+				const startMs = anchorCpuMs + Number(times[entry.beginIndex] - anchorGpuNs) / 1e6;
+				const endMs = anchorCpuMs + Number(times[entry.endIndex] - anchorGpuNs) / 1e6;
+				require_backend.emitTrace("webgpu", entry, startMs, endMs);
+			}
+		} finally {
+			batch.dst.unmap();
+			batch.querySet.destroy();
+			batch.resolve.destroy();
+			batch.dst.destroy();
+		}
+	});
+}
+
 //#endregion
 //#region src/backend/webgpu.ts
 /** Implementation of `Backend` that uses WebGPU in browsers. */
@@ -857,7 +987,7 @@ var WebGPUBackend = class {
 	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);
+		pipelineSubmit(this.device, exe, inputBuffers, outputBuffers);
 	}
 	#getBuffer(slot) {
 		const buffer = this.buffers.get(slot);
@@ -995,8 +1125,16 @@ function pipelineSource(device, kernel) {
 			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.Cast) {
+				const srcTy = dtypeToWgsl(src[0].dtype);
+				const dstTy = dtypeToWgsl(dtype);
+				if (require_backend.isFloatDtype(src[0].dtype) && !(require_backend.isFloatDtype(dtype) || dtype === require_backend.DType.Bool)) {
+					const maxVal = maxValueWgsl(dtype);
+					const x = isGensym(a) ? a : gensym();
+					if (x !== a) emit(`let ${x}: ${srcTy} = ${require_backend.strip1(a)};`);
+					source = `select(${dstTy}(${x}), ${maxVal}, ${x} >= ${srcTy}(${maxVal}))`;
+				} else source = `${dstTy}(${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) {
@@ -1099,12 +1237,14 @@ function pipelineSource(device, kernel) {
 		passes: [{ grid: [gridX, gridY] }]
 	};
 }
-function pipelineSubmit(device, pipelines, inputs, outputs) {
+function pipelineSubmit(device, exe, inputs, outputs) {
+	const { data: pipelines, source } = exe;
 	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 slot = maybeAcquireTracingSlot(device);
 		const bindGroup = device.createBindGroup({
 			layout: pipeline.getBindGroupLayout(0),
 			entries: [...inputs.map((buffer, i) => ({
@@ -1134,13 +1274,18 @@ function pipelineSubmit(device, pipelines, inputs, outputs) {
 		}
 		for (let i = 0; i < filteredPasses.length; i++) {
 			const { grid } = filteredPasses[i];
-			const passEncoder = commandEncoder.beginComputePass();
+			const passEncoder = commandEncoder.beginComputePass({ timestampWrites: slot ? {
+				querySet: slot.batch.querySet,
+				beginningOfPassWriteIndex: i === 0 ? slot.beginIndex : void 0,
+				endOfPassWriteIndex: i === filteredPasses.length - 1 ? slot.endIndex : void 0
+			} : void 0 });
 			passEncoder.setPipeline(pipeline);
 			passEncoder.setBindGroup(0, bindGroup);
 			if (uniformBindGroup) passEncoder.setBindGroup(1, uniformBindGroup, [i * uniformAlignment]);
 			passEncoder.dispatchWorkgroups(grid[0], grid[1]);
 			passEncoder.end();
 		}
+		if (slot) recordTrace(device, slot, source, filteredPasses.length, shader.code);
 	}
 	device.queue.submit([commandEncoder.finish()]);
 }

package/dist/{webgpu-B96vzWGE.js → webgpu-v_W_-oKw.js}

@@ -1,4 +1,4 @@
-import { AluExp, AluGroup, AluOp, DEBUG, DType, Executable, FpHash, Routines, SlotError, UnsupportedOpError, UnsupportedRoutineError, findPow2, isFloatDtype, mapSetUnion, prod, range, strip1, tuneWebgpu } from "./backend-nEolvdLv.js";
+import { AluExp, AluGroup, AluOp, DEBUG, DType, Executable, FpHash, Routines, SlotError, UnsupportedOpError, UnsupportedRoutineError, emitTrace, findPow2, isFloatDtype, isTracing, mapSetUnion, onFlushTrace, prod, range, strip1, traceSourceInfo, tuneWebgpu } from "./backend-Ctqs8la1.js";
 
 //#region src/backend/webgpu/builtins.ts
 const threefrySrc = `
@@ -152,6 +152,7 @@ var SyncReader = class SyncReader {
 		this.device = device;
 	}
 	#init() {
+		if (typeof OffscreenCanvas === "undefined") throw new Error("OffscreenCanvas is not available in this environment, so you cannot read data from WebGPU synchronously. Consider using the async API.");
 		const makeCanvas = () => new OffscreenCanvas(SyncReader.width, SyncReader.height);
 		this.deviceStorage = SyncReader.alphaModes.map(makeCanvas);
 		this.deviceContexts = this.deviceStorage.map((canvas, i) => {
@@ -247,6 +248,10 @@ function bitonicSortUniform(pass) {
 *   `2^(step+1)` with multiple workgroups. This doesn't use shared memory.
 *
 * The total number of passes is roughly `log2(n / workgroupSize)^2 / 2`.
+*
+* If `outputIndices` is true, the shader also tracks the original indices of
+* the sorted elements (argsort) and outputs them to a separate buffer. This
+* also makes the sorting algorithm stable.
 */
 function bitonicSortShader(device, dtype, n, batches, outputIndices) {
 	const ty = dtypeToWgsl(dtype, true);
@@ -286,14 +291,21 @@ ${isFloatDtype(dtype) ? `
 fn compare_and_swap(i: u32, j: u32) {
   let val_i = shared_vals[i];
   let val_j = shared_vals[j];
-  if (compare(val_j, val_i)) {
+${outputIndices ? `
+  if (
+    compare(val_j, val_i) ||
+    (!compare(val_i, val_j) && shared_idx[j] < shared_idx[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;` : ""}
-  }
+    shared_idx[j] = tmp_idx;
+  }` : `
+  if (compare(val_j, val_i)) {
+    shared_vals[i] = val_j;
+    shared_vals[j] = val_i;
+  }`}
 }
 
 @compute @workgroup_size(${workgroupSize})
@@ -370,13 +382,17 @@ ${outputIndices ? `
     if (j < ${n}u) {
       let val_i = output[base + i];
       let val_j = output[base + j];
-      if (compare(val_j, val_i)) {
+${outputIndices ? `
+      let idx_i = output_idx[base + i];
+      let idx_j = output_idx[base + j];
+      if (compare(val_j, val_i) || (!compare(val_i, val_j) && idx_j < idx_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;` : ""}
+        output_idx[base + i] = idx_j;
+        output_idx[base + j] = idx_i;` : `
+      if (compare(val_j, val_i)) {
+        output[base + i] = val_j;
+        output[base + j] = val_i;`}
       }
     }
   }
@@ -713,6 +729,120 @@ function createRoutineShader(device, routine) {
 	}
 }
 
+//#endregion
+//#region src/backend/webgpu/tracing.ts
+const MAX_TIMESTAMP_QUERIES = 4096;
+const activeBatch = /* @__PURE__ */ new WeakMap();
+function createTracingBatch(device) {
+	return {
+		querySet: device.createQuerySet({
+			type: "timestamp",
+			count: MAX_TIMESTAMP_QUERIES
+		}),
+		resolve: device.createBuffer({
+			size: MAX_TIMESTAMP_QUERIES * 8,
+			usage: GPUBufferUsage.QUERY_RESOLVE | GPUBufferUsage.COPY_SRC
+		}),
+		dst: device.createBuffer({
+			size: MAX_TIMESTAMP_QUERIES * 8,
+			usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST
+		}),
+		nextIndex: 0,
+		entries: []
+	};
+}
+function acquireTracingSlot(device) {
+	if (!device.features.has("timestamp-query")) return void 0;
+	let batch = activeBatch.get(device);
+	if (batch && batch.nextIndex >= MAX_TIMESTAMP_QUERIES) {
+		flushTracingBatch(device, batch);
+		batch = void 0;
+	}
+	if (!batch) {
+		batch = createTracingBatch(device);
+		activeBatch.set(device, batch);
+		onFlushTrace(() => {
+			const b = activeBatch.get(device);
+			if (b && b.entries.length > 0) flushTracingBatch(device, b);
+			activeBatch.delete(device);
+		});
+	}
+	const beginIndex = batch.nextIndex;
+	const endIndex = beginIndex + 1;
+	batch.nextIndex += 2;
+	return {
+		batch,
+		beginIndex,
+		endIndex
+	};
+}
+/**
+* If tracing is active, acquire a slot for timestamp queries.
+*
+* Returns undefined if tracing is not active or the device doesn't support
+* timestamp queries.
+*/
+function maybeAcquireTracingSlot(device) {
+	if (!isTracing()) return void 0;
+	return acquireTracingSlot(device);
+}
+/**
+* Record a tracing entry for a pipeline dispatch and schedule an auto-flush.
+*/
+function recordTrace(device, slot, source, numPasses, wgslSource) {
+	const info = traceSourceInfo(source);
+	info.properties.push(["passes", `${numPasses}`]);
+	info.properties.push(["source", wgslSource]);
+	slot.batch.entries.push({
+		...info,
+		beginIndex: slot.beginIndex,
+		endIndex: slot.endIndex
+	});
+	scheduleAutoFlush(device);
+}
+/**
+* If the active batch has pending entries, flush and replace it so traces
+* are emitted without waiting for the batch to fill or stopTrace().
+*
+* Called after each dispatch records its entry via a microtask so that
+* synchronous back-to-back dispatches are still batched together.
+*/
+function scheduleAutoFlush(device) {
+	queueMicrotask(() => {
+		const batch = activeBatch.get(device);
+		if (batch && batch.entries.length > 0) {
+			flushTracingBatch(device, batch);
+			activeBatch.set(device, createTracingBatch(device));
+		}
+	});
+}
+function flushTracingBatch(device, batch) {
+	if (batch.entries.length === 0) return;
+	const usedQueries = batch.nextIndex;
+	const encoder = device.createCommandEncoder();
+	encoder.resolveQuerySet(batch.querySet, 0, usedQueries, batch.resolve, 0);
+	encoder.copyBufferToBuffer(batch.resolve, 0, batch.dst, 0, usedQueries * 8);
+	device.queue.submit([encoder.finish()]);
+	const { entries } = batch;
+	batch.dst.mapAsync(GPUMapMode.READ).then(() => {
+		try {
+			const times = new BigInt64Array(batch.dst.getMappedRange());
+			const anchorGpuNs = times[entries[entries.length - 1].endIndex];
+			const anchorCpuMs = performance.now();
+			for (const entry of entries) {
+				const startMs = anchorCpuMs + Number(times[entry.beginIndex] - anchorGpuNs) / 1e6;
+				const endMs = anchorCpuMs + Number(times[entry.endIndex] - anchorGpuNs) / 1e6;
+				emitTrace("webgpu", entry, startMs, endMs);
+			}
+		} finally {
+			batch.dst.unmap();
+			batch.querySet.destroy();
+			batch.resolve.destroy();
+			batch.dst.destroy();
+		}
+	});
+}
+
 //#endregion
 //#region src/backend/webgpu.ts
 /** Implementation of `Backend` that uses WebGPU in browsers. */
@@ -857,7 +987,7 @@ var WebGPUBackend = class {
 	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);
+		pipelineSubmit(this.device, exe, inputBuffers, outputBuffers);
 	}
 	#getBuffer(slot) {
 		const buffer = this.buffers.get(slot);
@@ -995,8 +1125,16 @@ function pipelineSource(device, kernel) {
 			else if (op === AluOp.Reciprocal) source = `(1.0 / ${a})`;
 			else if (op === AluOp.Floor) source = `floor(${strip1(a)})`;
 			else if (op === AluOp.Ceil) source = `ceil(${strip1(a)})`;
-			else if (op === AluOp.Cast) source = `${dtypeToWgsl(dtype)}(${strip1(a)})`;
-			else if (op === AluOp.Bitcast) source = `bitcast<${dtypeToWgsl(dtype)}>(${strip1(a)})`;
+			else if (op === AluOp.Cast) {
+				const srcTy = dtypeToWgsl(src[0].dtype);
+				const dstTy = dtypeToWgsl(dtype);
+				if (isFloatDtype(src[0].dtype) && !(isFloatDtype(dtype) || dtype === DType.Bool)) {
+					const maxVal = maxValueWgsl(dtype);
+					const x = isGensym(a) ? a : gensym();
+					if (x !== a) emit(`let ${x}: ${srcTy} = ${strip1(a)};`);
+					source = `select(${dstTy}(${x}), ${maxVal}, ${x} >= ${srcTy}(${maxVal}))`;
+				} else source = `${dstTy}(${strip1(a)})`;
+			} else if (op === AluOp.Bitcast) source = `bitcast<${dtypeToWgsl(dtype)}>(${strip1(a)})`;
 		}
 		else if (op === AluOp.Where) source = `select(${strip1(gen(src[2]))}, ${strip1(gen(src[1]))}, ${strip1(gen(src[0]))})`;
 		else if (op === AluOp.Threefry2x32) {
@@ -1099,12 +1237,14 @@ function pipelineSource(device, kernel) {
 		passes: [{ grid: [gridX, gridY] }]
 	};
 }
-function pipelineSubmit(device, pipelines, inputs, outputs) {
+function pipelineSubmit(device, exe, inputs, outputs) {
+	const { data: pipelines, source } = exe;
 	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 }) => prod(grid) > 0);
 		if (filteredPasses.length === 0) continue;
+		const slot = maybeAcquireTracingSlot(device);
 		const bindGroup = device.createBindGroup({
 			layout: pipeline.getBindGroupLayout(0),
 			entries: [...inputs.map((buffer, i) => ({
@@ -1134,13 +1274,18 @@ function pipelineSubmit(device, pipelines, inputs, outputs) {
 		}
 		for (let i = 0; i < filteredPasses.length; i++) {
 			const { grid } = filteredPasses[i];
-			const passEncoder = commandEncoder.beginComputePass();
+			const passEncoder = commandEncoder.beginComputePass({ timestampWrites: slot ? {
+				querySet: slot.batch.querySet,
+				beginningOfPassWriteIndex: i === 0 ? slot.beginIndex : void 0,
+				endOfPassWriteIndex: i === filteredPasses.length - 1 ? slot.endIndex : void 0
+			} : void 0 });
 			passEncoder.setPipeline(pipeline);
 			passEncoder.setBindGroup(0, bindGroup);
 			if (uniformBindGroup) passEncoder.setBindGroup(1, uniformBindGroup, [i * uniformAlignment]);
 			passEncoder.dispatchWorkgroups(grid[0], grid[1]);
 			passEncoder.end();
 		}
+		if (slot) recordTrace(device, slot, source, filteredPasses.length, shader.code);
 	}
 	device.queue.submit([commandEncoder.finish()]);
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@jax-js/jax",
-  "version": "0.1.8",
+  "version": "0.1.10",
   "description": "Numerical computing and ML in the browser",
   "keywords": [
     "machine learning",
@@ -38,15 +38,13 @@
   "devDependencies": {
     "@eslint/js": "^9.31.0",
     "@types/debug": "^4.1.12",
-    "@vitest/browser-playwright": "^4.0.9",
-    "@vitest/coverage-v8": "4.0.9",
+    "@vitest/browser-playwright": "^4.1.0",
+    "@vitest/coverage-v8": "^4.1.0",
     "@webgpu/types": "^0.1.68",
     "eslint": "^9.31.0",
     "eslint-plugin-import": "^2.32.0",
     "globals": "^16.0.0",
-    "husky": "^9.1.7",
-    "lint-staged": "^16.2.7",
-    "playwright": "~1.52.0",
+    "playwright": "~1.58.2",
     "prettier": "^3.6.2",
     "prettier-plugin-svelte": "^3.4.0",
     "tsdown": "^0.13.2",
@@ -55,7 +53,7 @@
     "typedoc-theme-fresh": "^0.2.3",
     "typescript": "~5.9.3",
     "typescript-eslint": "^8.46.4",
-    "vitest": "^4.0.9"
+    "vitest": "^4.1.0"
   },
   "engines": {
     "pnpm": ">=10.0.0"
@@ -76,15 +74,6 @@
     ],
     "proseWrap": "always"
   },
-  "lint-staged": {
-    "*.{ts,tsx,js,jsx}": [
-      "eslint --fix",
-      "prettier --write"
-    ],
-    "*.{json,md,yml,yaml,css,svelte,html}": [
-      "prettier --write"
-    ]
-  },
   "scripts": {
     "build": "tsdown",
     "build:watch": "TSDOWN_WATCH_MODE=1 tsdown",