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 2 commits
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
208 changes: 170 additions & 38 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 +3,24 @@

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, ProgramInfo, ProgramInputTensorInfoDependency, QueryType} from './webgpu/types';

interface KernelInfo {
gyagp marked this conversation as resolved.
Show resolved Hide resolved
opType: string;
nodeName: string;
kernelEntry: RunFunction;
attributes: [((attribute: unknown) => unknown)|undefined, unknown];
programName: string;
inputTensorViews: readonly TensorView[];
outputTensorViews: readonly TensorView[];
gyagp marked this conversation as resolved.
Show resolved Hide resolved
}
gyagp marked this conversation as resolved.
Show resolved Hide resolved

const getProgramInputTensorInfoDependencyKey =
(inputTensors: readonly TensorView[], inputDependencies: readonly ProgramInputTensorInfoDependency[]): string => {
Expand Down Expand Up @@ -122,20 +134,21 @@ export class WebGpuBackend {
return data;
}

/**
* a KernelID -> kernel info mapping. value is
* [ op_type, name, run function, [optional] preprocess_attribute_once function ]
*/
kernels: Map<number, [string, string, RunFunction, [((attribute: unknown) => unknown) | undefined, unknown]]>;

kernels: Map<number, KernelInfo>;
gyagp marked this conversation as resolved.
Show resolved Hide resolved
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
private pendingKernels: number[] = [];
// queryReadData -> pendingKernels mapping for all the batches
private pendingQueries: Map<number, number[]> = new Map();
private queryResolveData?: GpuData;
private querySet?: GPUQuerySet;
private queryTimeBase?: bigint;
queryType: QueryType;

env: Env;

Expand All @@ -161,7 +174,9 @@ 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 All @@ -188,36 +203,47 @@ export class WebGpuBackend {
};

Object.defineProperty(this.env.webgpu, 'device', {value: this.device});

// init queryType, which is necessary for createKernel
this.setQueryType();
}

dispose(): void {
if (typeof this.querySet !== 'undefined') {
this.querySet.destroy();
this.gpuDataManager.release(this.queryResolveData!.id);
gyagp marked this conversation as resolved.
Show resolved Hide resolved
}
this.gpuDataManager.dispose();
}

getCommandEncoder(): GPUCommandEncoder {
if (!this.commandEncoder) {
this.commandEncoder = this.device.createCommandEncoder();

// refresh queryType, as sometimes we only need to enable query for a specific run
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 +260,87 @@ 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];
const kernelInfo = this.kernels.get(kernelId)!;
const opType = kernelInfo.opType;
const nodeName = kernelInfo.nodeName;
const programName = kernelInfo.programName;
const inputTensorViews = kernelInfo.inputTensorViews;
const outputTensorViews = kernelInfo.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: opType,
kernelName: nodeName,
gyagp marked this conversation as resolved.
Show resolved Hide resolved
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}|${opType}|${nodeName}|${programName}" ${inputShapes}${
outputShapes}execution time: ${endTime - startTime} ns`);
}
}
queryReadData.buffer.unmap();
this.gpuDataManager.release(queryReadData.id);
gyagp marked this conversation as resolved.
Show resolved Hide resolved
this.pendingQueries.delete(queryReadData.id);
});
}
}

/**
Expand Down Expand Up @@ -378,14 +472,22 @@ export class WebGpuBackend {
this.programManager.setArtifact(key, artifact);
LOG_DEBUG('info', () => `[artifact] key: ${key}, programName: ${program.name}`);
}
// update kernels
const kernelInfo = this.kernels.get(this.currentKernelId!)!;
gyagp marked this conversation as resolved.
Show resolved Hide resolved
kernelInfo.programName = artifact.programInfo.name;
kernelInfo.inputTensorViews = inputTensorViews;
kernelInfo.outputTensorViews = outputTensorViews;

LOG_DEBUG(
'info',
() => `[ProgramManager] run "${program.name}" (key=${key}) with ${normalizedDispatchGroup[0]}x${
normalizedDispatchGroup[1]}x${normalizedDispatchGroup[2]}`);
this.programManager.run(
artifact, inputTensorViews, outputTensorViews, inputDatas, outputDatas, normalizedDispatchGroup,
uniformBufferBinding);

if (this.queryType !== QueryType.none) {
this.pendingKernels.push(this.currentKernelId!);
}

this.programManager.run(artifact, inputDatas, outputDatas, normalizedDispatchGroup, uniformBufferBinding);

return outputTensorViews;
}
Expand Down Expand Up @@ -418,7 +520,16 @@ export class WebGpuBackend {
throw new Error(`kernel not implemented: ${opType}`);
}

this.kernels.set(kernelId, [opType, nodeName, op[0], [op[1], attribute]]);
const kernelInfo: KernelInfo = {
opType,
nodeName,
kernelEntry: op[0],
attributes: [op[1], attribute],
programName: '',
inputTensorViews: [],
outputTensorViews: [],
};
this.kernels.set(kernelId, kernelInfo);
}

releaseKernel(kernelId: number): void {
Expand All @@ -439,7 +550,10 @@ export class WebGpuBackend {
if (!kernel) {
throw new Error(`kernel not created: ${kernelId}`);
}
const [opType, nodeName, kernelEntry, attributes] = kernel;
const opType = kernel.opType;
const nodeName = kernel.nodeName;
const kernelEntry = kernel.kernelEntry;
const attributes = kernel.attributes;
if (this.currentKernelId !== null) {
throw new Error(`kernel "[${opType}] ${nodeName}" is not allowed to be called recursively`);
}
Expand Down Expand Up @@ -514,5 +628,23 @@ 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
}
9 changes: 5 additions & 4 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 @@ -186,9 +186,10 @@ 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)) : `${kernel}`),
(opType: string, kernelId: number, attribute: unknown) => backend.createKernel(
opType, kernelId, attribute,
env.debug || backend.queryType !== QueryType.none ? module.UTF8ToString(module._JsepGetNodeName(kernelId)) :
`${opType}`),
gyagp marked this conversation as resolved.
Show resolved Hide resolved

// jsepReleaseKernel
(kernel: number) => backend.releaseKernel(kernel),
Expand Down
Loading