Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[js/webgpu] Refactor timestamp-query and introduce timestamp-query-inside-passes #18894

Merged
merged 11 commits into from
Jan 13, 2024
Merged

[js/webgpu] Refactor timestamp-query and introduce timestamp-query-inside-passes #18894

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
4ee5b41
[js/webgpu] Experiment on timestamp query between atPasses and inside…
Dec 20, 2023
4745182
Address Jiajia's comments and refactor kernels
Dec 27, 2023
c4951a1
Add more comments from Jiajia
Dec 27, 2023
585ec83
Address Yulong's comments
Jan 3, 2024
40790b2
Unify the terms
Jan 4, 2024
3504a34
Incorporate trace support
Jan 4, 2024
e0859ee
Address Yulong's comments
Jan 11, 2024
982a99e
Put TimestampQuery to the end of file
Jan 11, 2024
bbe09ad
Merge branch 'main' of https://github.com/gyagp/onnxruntime into time…
Jan 11, 2024
ab20d47
Call endComputePass in flush and fix the name of writeTimestamp
Jan 12, 2024
bfdf6aa
Merge branch 'main' of https://github.com/gyagp/onnxruntime into time…
Jan 13, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
155 changes: 128 additions & 27 deletions js/web/lib/wasm/jsep/backend-webgpu.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,12 +3,14 @@

import {Env, Tensor} from 'onnxruntime-common';

import {tensorDataTypeEnumToString} from '../wasm-common';

import {configureLogger, LOG_DEBUG} from './log';
import {createView, TensorView} from './tensor-view';
import {createGpuDataManager, downloadGpuData, GpuDataManager} from './webgpu/gpu-data-manager';
import {RunFunction, WEBGPU_OP_RESOLVE_RULES} from './webgpu/op-resolve-rules';
import {ProgramManager} from './webgpu/program-manager';
import {ComputeContext, GpuData, ProgramInfo, ProgramInputTensorInfoDependency} from './webgpu/types';
import {ComputeContext, GpuData, PendingKernelInfo, ProgramInfo, ProgramInputTensorInfoDependency, QueryType} from './webgpu/types';

const getProgramInputTensorInfoDependencyKey =
(inputTensors: readonly TensorView[], inputDependencies: readonly ProgramInputTensorInfoDependency[]): string => {
Expand Down Expand Up @@ -130,12 +132,18 @@ export class WebGpuBackend {

private commandEncoder: GPUCommandEncoder|null = null;
private computePassEncoder: GPUComputePassEncoder|null = null;
maxDispatchNumber = 16;
pendingDispatchNumber = 0;

queryData?: GpuData;
querySet?: GPUQuerySet;
querySetCount = 2;
queryTimeBase?: bigint;

// info of kernels pending submission for a single batch
pendingKernels: PendingKernelInfo[] = [];
// queryReadData -> pendingKernels mapping for all the batches
private pendingQueries: Map<number, PendingKernelInfo[]> = new Map();
private queryResolveData?: GpuData;
private querySet?: GPUQuerySet;
private queryTimeBase?: bigint;
queryType: QueryType;

env: Env;

Expand All @@ -161,7 +169,9 @@ export class WebGpuBackend {
requiredFeatures,
};

if (adapter.features.has('timestamp-query')) {
if (adapter.features.has('chromium-experimental-timestamp-query-inside-passes')) {
requiredFeatures.push('chromium-experimental-timestamp-query-inside-passes' as GPUFeatureName);
} else if (adapter.features.has('timestamp-query')) {
requiredFeatures.push('timestamp-query');
}
if (adapter.features.has('shader-f16')) {
Expand Down Expand Up @@ -200,24 +210,30 @@ export class WebGpuBackend {
getCommandEncoder(): GPUCommandEncoder {
if (!this.commandEncoder) {
this.commandEncoder = this.device.createCommandEncoder();

this.setQueryType();
if (this.queryType !== QueryType.none && typeof this.querySet === 'undefined') {
this.querySet = this.device.createQuerySet({
type: 'timestamp',
count: this.maxDispatchNumber * 2,
});
this.queryResolveData = this.gpuDataManager.create(
// eslint-disable-next-line no-bitwise
this.maxDispatchNumber * 2 * 8, GPUBufferUsage.COPY_SRC | GPUBufferUsage.QUERY_RESOLVE);
}
}
return this.commandEncoder;
}

getComputePassEncoder(): GPUComputePassEncoder {
if (!this.computePassEncoder) {
const computePassDescriptor: GPUComputePassDescriptor = {};
if (this.isQueryEnabled()) {
if (typeof this.querySet === 'undefined') {
this.querySet = this.device.createQuerySet({
type: 'timestamp',
count: this.querySetCount,
});
}

if (this.queryType === QueryType.atPasses) {
computePassDescriptor.timestampWrites = {
querySet: this.querySet,
beginningOfPassWriteIndex: 0,
endOfPassWriteIndex: 1,
querySet: this.querySet!,
beginningOfPassWriteIndex: this.pendingDispatchNumber * 2,
endOfPassWriteIndex: this.pendingDispatchNumber * 2 + 1,
};
}

Expand All @@ -234,19 +250,85 @@ export class WebGpuBackend {
}

flush(): void {
if (this.commandEncoder) {
this.endComputePass();
this.device.queue.submit([this.getCommandEncoder().finish()]);
this.gpuDataManager.refreshPendingBuffers();
this.commandEncoder = null;
this.pendingDispatchNumber = 0;
if (!this.commandEncoder) {
return;
}
}

isQueryEnabled(): boolean {
return this.device.features.has('timestamp-query') &&
(this.env.webgpu.profiling?.mode === 'default' ||
(!this.env.webgpu.profiling?.mode && this.env.webgpu.profilingMode === 'default'));
let queryReadData: GpuData;
if (this.queryType !== QueryType.none) {
this.commandEncoder.resolveQuerySet(
this.querySet!, 0, this.pendingDispatchNumber * 2, this.queryResolveData!.buffer, 0);
queryReadData = this.gpuDataManager.create(
gyagp marked this conversation as resolved.
Show resolved Hide resolved
// eslint-disable-next-line no-bitwise
this.pendingDispatchNumber * 2 * 8, GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST);
this.pendingQueries.set(queryReadData.id, this.pendingKernels);
this.pendingKernels = [];
this.commandEncoder.copyBufferToBuffer(
this.queryResolveData!.buffer, 0, queryReadData.buffer, 0, this.pendingDispatchNumber * 2 * 8);
}

this.device.queue.submit([this.commandEncoder.finish()]);
this.gpuDataManager.refreshPendingBuffers();
this.commandEncoder = null;
this.pendingDispatchNumber = 0;

if (this.queryType !== QueryType.none) {
void queryReadData!.buffer.mapAsync(GPUMapMode.READ).then(() => {
const mappedData = new BigUint64Array(queryReadData.buffer.getMappedRange());
const pendingKernels = this.pendingQueries.get(queryReadData.id);
for (let i = 0; i < mappedData.length / 2; i++) {
const kernelId = pendingKernels![i].id;
const kernelInfo = this.kernels.get(kernelId)!;
const kernelType = kernelInfo[0];
const kernelName = pendingKernels![i].name;
const inputTensorViews = pendingKernels![i].inputTensorViews;
const outputTensorViews = pendingKernels![i].outputTensorViews;
const startTimeU64 = mappedData[i * 2];
const endTimeU64 = mappedData[i * 2 + 1];

if (typeof this.queryTimeBase === 'undefined') {
this.queryTimeBase = startTimeU64;
}

const startTime = Number(startTimeU64 - this.queryTimeBase);
const endTime = Number(endTimeU64 - this.queryTimeBase);

if (!Number.isSafeInteger(startTime) || !Number.isSafeInteger(endTime)) {
throw new RangeError('incorrect timestamp range');
}

if (this.env.webgpu.profiling?.ondata) {
this.env.webgpu.profiling.ondata({
version: 1,
inputsMetadata: inputTensorViews.map(
value => ({dims: value.dims, dataType: tensorDataTypeEnumToString(value.dataType)})),
outputsMetadata: outputTensorViews.map(
value => ({dims: value.dims, dataType: tensorDataTypeEnumToString(value.dataType)})),
kernelId,
kernelType,
kernelName,
startTime,
endTime,
});
} else {
// if no callback is provided, print the profiling message to console
let inputShapes = '';
inputTensorViews.forEach((value, i) => {
inputShapes += `input[${i}]: [${value.dims}] | ${tensorDataTypeEnumToString(value.dataType)}, `;
});
let outputShapes = '';
outputTensorViews.forEach((value, i) => {
outputShapes += `output[${i}]: [${value.dims}] | ${tensorDataTypeEnumToString(value.dataType)}, `;
});
// eslint-disable-next-line no-console
console.log(`[profiling] kernel "${kernelId}|${kernelName}" ${inputShapes}${outputShapes}execution time: ${
endTime - startTime} ns`);
}
}
queryReadData.buffer.unmap();
this.gpuDataManager.release(queryReadData.id);
gyagp marked this conversation as resolved.
Show resolved Hide resolved
});
}
}

/**
Expand Down Expand Up @@ -514,5 +596,24 @@ export class WebGpuBackend {
return createView(data.buffer, type);
};
}
writeTimeStamp(index: number): void {
if (this.queryType !== QueryType.insidePasses) {
return;
}

// eslint-disable-next-line @typescript-eslint/no-explicit-any
(this.computePassEncoder as any).writeTimestamp(this.querySet, index);
}

setQueryType(): void {
this.queryType = QueryType.none;
if (this.env.webgpu.profiling?.mode === 'default') {
gyagp marked this conversation as resolved.
Show resolved Hide resolved
if (this.device.features.has('chromium-experimental-timestamp-query-inside-passes')) {
this.queryType = QueryType.insidePasses;
} else if (this.device.features.has('timestamp-query')) {
this.queryType = QueryType.atPasses;
}
gyagp marked this conversation as resolved.
Show resolved Hide resolved
}
}
// #endregion
}
5 changes: 3 additions & 2 deletions js/web/lib/wasm/jsep/init.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,7 @@ import {WebGpuBackend} from './backend-webgpu';
import {LOG_DEBUG} from './log';
import {TensorView} from './tensor-view';
import {ShapeUtil} from './util';
import {ComputeContext, ComputeContextInputsOutputsMapping, ProgramInfo} from './webgpu/types';
import {ComputeContext, ComputeContextInputsOutputsMapping, ProgramInfo, QueryType} from './webgpu/types';

/* eslint-disable no-bitwise */

Expand Down Expand Up @@ -188,7 +188,8 @@ export const init = async(module: OrtWasmModule, env: Env, gpuAdapter: GPUAdapte
// jsepCreateKernel
(name: string, kernel: number, attribute: unknown) => backend.createKernel(
name, kernel, attribute,
env.debug || backend.isQueryEnabled() ? module.UTF8ToString(module._JsepGetNodeName(kernel)) : `${kernel}`),
env.debug || backend.queryType !== QueryType.none ? module.UTF8ToString(module._JsepGetNodeName(kernel)) :
`${kernel}`),

// jsepReleaseKernel
(kernel: number) => backend.releaseKernel(kernel),
Expand Down
91 changes: 21 additions & 70 deletions js/web/lib/wasm/jsep/webgpu/program-manager.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,13 +1,12 @@
// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.

import {tensorDataTypeEnumToString} from '../../wasm-common';
import {WebGpuBackend} from '../backend-webgpu';
import {LOG_DEBUG} from '../log';
import {TensorView} from '../tensor-view';

import {createShaderHelper} from './ops/common';
import {Artifact, GpuData, ProgramInfo} from './types';
import {Artifact, GpuData, PendingKernelInfo, ProgramInfo, QueryType} from './types';

/**
* ProgramManager is the main class behind running computations
Expand Down Expand Up @@ -36,8 +35,8 @@ export class ProgramManager {
inputs: GpuData[], outputs: GpuData[], dispatchGroup: [number, number, number],
uniformBufferBinding: GPUBindingResource|undefined): void {
const device = this.backend.device;

const computePassEncoder = this.backend.getComputePassEncoder();
this.backend.writeTimeStamp(this.backend.pendingDispatchNumber * 2);
computePassEncoder.setPipeline(buildArtifact.computePipeline);
const entries = [];
for (const input of inputs) {
Expand All @@ -55,77 +54,29 @@ export class ProgramManager {

computePassEncoder.dispatchWorkgroups(...dispatchGroup);

this.backend.pendingDispatchNumber++;

if (this.backend.isQueryEnabled()) {
if (typeof this.backend.queryData === 'undefined') {
this.backend.queryData = this.backend.gpuDataManager.create(
// eslint-disable-next-line no-bitwise
this.backend.querySetCount * 8, GPUBufferUsage.COPY_SRC | GPUBufferUsage.QUERY_RESOLVE);
}
const syncData = this.backend.gpuDataManager.create(
// eslint-disable-next-line no-bitwise
this.backend.querySetCount * 8, GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST);

this.backend.endComputePass();
this.backend.getCommandEncoder().resolveQuerySet(this.backend.querySet!, 0, 2, this.backend.queryData.buffer, 0);
this.backend.getCommandEncoder().copyBufferToBuffer(
this.backend.queryData.buffer, 0, syncData.buffer, 0, this.backend.querySetCount * 8);
this.backend.flush();

if (this.backend.queryType !== QueryType.none) {
const kernelId = this.backend.currentKernelId!;
const kernelInfo = this.backend.kernels.get(kernelId)!;

void syncData.buffer.mapAsync(GPUMapMode.READ).then(() => {
const mappedData = new BigUint64Array(syncData.buffer.getMappedRange());
const [startTimeU64, endTimeU64] = mappedData;
const [kernelType, kernelName] = kernelInfo;

syncData.buffer.unmap();

if (typeof this.backend.queryTimeBase === 'undefined') {
this.backend.queryTimeBase = startTimeU64;
}

const startTime = Number(startTimeU64 - this.backend.queryTimeBase);
const endTime = Number(endTimeU64 - this.backend.queryTimeBase);

if (!Number.isSafeInteger(startTime) || !Number.isSafeInteger(endTime)) {
throw new RangeError('incorrect timestamp range');
}

this.backend.gpuDataManager.release(syncData.id);
if (this.backend.env.webgpu.profiling?.ondata) {
this.backend.env.webgpu.profiling.ondata({
version: 1,
inputsMetadata: inputTensorViews.map(
value => ({dims: value.dims, dataType: tensorDataTypeEnumToString(value.dataType)})),
outputsMetadata: outputTensorViews.map(
value => ({dims: value.dims, dataType: tensorDataTypeEnumToString(value.dataType)})),
kernelId,
kernelType,
kernelName,
startTime,
endTime,
});
} else {
// if no callback is provided, print the profiling message to console
let inputShapes = '';
inputTensorViews.forEach((value, i) => {
inputShapes += `input[${i}]: [${value.dims}] | ${tensorDataTypeEnumToString(value.dataType)}, `;
});
let outputShapes = '';
outputTensorViews.forEach((value, i) => {
outputShapes += `output[${i}]: [${value.dims}] | ${tensorDataTypeEnumToString(value.dataType)}, `;
});
// eslint-disable-next-line no-console
console.log(`[profiling] kernel "${kernelId}|${kernelName}|${buildArtifact.programInfo.name}" ${inputShapes}${
outputShapes}execution time: ${endTime - startTime} ns`);
}
});
let kernelName = kernelInfo[0];
if (buildArtifact.programInfo.name !== kernelName) {
kernelName = `${kernelName}/${buildArtifact.programInfo.name}`;
gyagp marked this conversation as resolved.
Show resolved Hide resolved
}
const pendingKernelInfo: PendingKernelInfo = {
id: kernelId,
name: kernelName,
inputTensorViews,
outputTensorViews,
};
this.backend.pendingKernels.push(pendingKernelInfo);
gyagp marked this conversation as resolved.
Show resolved Hide resolved
this.backend.writeTimeStamp(this.backend.pendingDispatchNumber * 2 + 1);
}

if (this.backend.pendingDispatchNumber >= 16) {
this.backend.pendingDispatchNumber++;
if (this.backend.pendingDispatchNumber >= this.backend.maxDispatchNumber ||
this.backend.queryType === QueryType.atPasses) {
this.backend.endComputePass();
}
if (this.backend.pendingDispatchNumber >= this.backend.maxDispatchNumber) {
this.backend.flush();
}
}
Expand Down
14 changes: 13 additions & 1 deletion js/web/lib/wasm/jsep/webgpu/types.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,12 @@ export enum GpuDataType {
}
export type GpuDataId = number;

export enum QueryType {
none,
insidePasses,
atPasses,
}
gyagp marked this conversation as resolved.
Show resolved Hide resolved

export interface GpuData {
type: GpuDataType;
id: GpuDataId;
Expand All @@ -23,12 +29,18 @@ export interface TensorInfo {
dataType: number;
}


export interface ProgramUniform {
type: 'int32'|'float32'|'uint32';
data: number|readonly number[];
}

export interface PendingKernelInfo {
id: number;
name: string;
inputTensorViews: readonly TensorView[];
outputTensorViews: readonly TensorView[];
}

/**
* Represent the dependency of a program on a specific input tensor.
*
Expand Down