mirror of
https://github.com/saymrwulf/onnxruntime.git
synced 2026-06-03 23:49:44 +00:00
[js/webgpu] Optimize NCHW layout for InstanceNormalization (#18123)
### Description The changes in this PR includes: 1) Fix f16 errors in InstanceNormalization with NCHW format. 2) Use vec to further optimize the original algorithm. 3) (Removed) Don't do layout conversion for InstanceNormalization for JSEP since InstanceNormalization itself is suitable for NCHW layout and has better performance in our current implementation. Tested on sd-vae-decoder-f16.onnx, it becomes 285 ms from 314 ms. The aggregate gpu profiling data can be found as below (Note the data is based change 3).): Before: <html> <body> <!--StartFragment--><span><span class="ui-provider ef bbg bbh bbi bbj bbk bbl bbm bbn bbo bbp bbq bbr bbs bbt bbu bbv bbw bbx bby bbz bca bcb bcc bcd bce bcf bcg bch bci bcj bck bcl bcm bcn" dir="ltr"> Kernel | Time (Ms) | Percentage (%) -- | -- | -- Conv | 201.55 | 69.56 InstanceNormalization | 42.49 | 14.67 Transpose | 28.95 | 9.99 Mul | 5.69 | 1.96 Add | 3.82 | 1.32 MatMul | 3.27 | 1.13 Sigmoid | 2.24 | 0.77 Resize | 1.16 | 0.40 Softmax | 0.34 | 0.12 Cast | 0.24 | 0.08 Sum | 289.75 <br class="Apple-interchange-newline"><!--EndFragment--> </body> </html> After: <html> <body> <!--StartFragment--><span><span class="ui-provider ef bbg bbh bbi bbj bbk bbl bbm bbn bbo bbp bbq bbr bbs bbt bbu bbv bbw bbx bby bbz bca bcb bcc bcd bce bcf bcg bch bci bcj bck bcl bcm bcn" dir="ltr"> Kernel | Time (Ms) | Percentage (%) -- | -- | -- Conv | 205.44 | 79.43 InstanceNormalization | 18.24 | 7.05 Transpose | 17.64 | 6.82 Mul | 5.69 | 2.20 Add | 3.81 | 1.47 MatMul | 3.56 | 1.38 Sigmoid | 2.24 | 0.86 Resize | 1.19 | 0.46 Softmax | 0.59 | 0.23 Cast | 0.24 | 0.09 Sum | 258.65 | </span></span><!--EndFragment--> </body> </html> From above table, we can see that two ops time are greatly reduced. One is InstanceNormalization and the other is Transpose. The reason that the transpose time is reduced is because each InstanceNormalization is surrounded with two reshape ops in sd-vae-decoder-f16.onnx. Due to JSEP is prefer NHWC and InstanceNormalization is layout sensitive op, so two extra transpose ops are inserted dynamically when executing this model. After this change, those inserted transpose ops are not needed anymore. So the overall transpose time is reduced.
This commit is contained in:
Jiajia Qin
GitHub
parent
4bbed4c71a
commit
8f7b89bd5b
1 changed files with 23 additions and 19 deletions
|
|
@ -7,7 +7,7 @@ import {ShapeUtil} from '../../util';
|
|||
import {AttributeWithCacheKey, createAttributeWithCacheKey} from '../attribute-with-cache-key';
|
||||
import {ComputeContext, ProgramInfo} from '../types';
|
||||
|
||||
import {fillVector, getMaxComponents, inputVariable, outputVariable, ShaderHelper, tensorTypeToWsglStorageType} from './common';
|
||||
import {fillVector, getMaxComponents, inputVariable, outputVariable, ShaderHelper, sumVector, tensorTypeToWsglStorageType} from './common';
|
||||
|
||||
export interface InstanceNormAttributes extends AttributeWithCacheKey {
|
||||
epsilon: number;
|
||||
|
|
@ -26,22 +26,25 @@ const createInstanceNormProgramInfo =
|
|||
const axis = 2;
|
||||
const normCount = ShapeUtil.sizeToDimension(xShape, axis);
|
||||
const normSize = ShapeUtil.sizeFromDimension(xShape, axis);
|
||||
const components = getMaxComponents(normSize);
|
||||
const normPackedSize = normSize / components;
|
||||
const C = xShape[1];
|
||||
const x = inputVariable('x', inputs[0].dataType, [xShape[0], xShape[1], normSize]);
|
||||
const x = inputVariable('x', inputs[0].dataType, [xShape[0], xShape[1], normPackedSize], components);
|
||||
const scale = inputVariable('scale', inputs[1].dataType, inputs[1].dims);
|
||||
const bias = inputVariable('bias', inputs[2].dataType, inputs[2].dims);
|
||||
const output = outputVariable('output', inputs[0].dataType, [xShape[0], xShape[1], normSize]);
|
||||
const output = outputVariable('output', inputs[0].dataType, [xShape[0], xShape[1], normPackedSize], components);
|
||||
const variables = [x, scale, bias, output];
|
||||
const dataType = x.type.value;
|
||||
const f32Type = components === 1 ? 'f32' : `vec${components}<f32>`;
|
||||
const workgroupSize = 64;
|
||||
const getShaderSource = (shaderHelper: ShaderHelper) => `
|
||||
|
||||
const C: u32 = ${C};
|
||||
const normSize: u32 = ${normSize};
|
||||
const epsilon: f32 = ${attributes.epsilon};
|
||||
var<workgroup> meanShared : ${dataType};
|
||||
var<workgroup> squaredNormShared : ${dataType};
|
||||
var<workgroup> workgroupShared : array<${dataType}, ${workgroupSize}>;
|
||||
var<workgroup> meanShared : f32;
|
||||
var<workgroup> squaredNormShared : f32;
|
||||
var<workgroup> workgroupShared : array<${f32Type}, ${workgroupSize}>;
|
||||
const workgroupSize = ${workgroupSize}u;
|
||||
${shaderHelper.declareVariables(...variables)}
|
||||
${shaderHelper.mainStart(workgroupSize)}
|
||||
|
|
@ -51,9 +54,9 @@ const createInstanceNormProgramInfo =
|
|||
let localIndex = local_id.x;
|
||||
|
||||
// initialize workgroup memory
|
||||
var initial: ${dataType} = 0;
|
||||
for (var h = localIndex; h < normSize; h += workgroupSize) {
|
||||
initial = initial + ${x.get('batch', 'channel', 'h')};
|
||||
var initial = ${f32Type}(0);
|
||||
for (var h = localIndex; h < ${normPackedSize}; h += workgroupSize) {
|
||||
initial = initial + ${f32Type}(${x.get('batch', 'channel', 'h')});
|
||||
}
|
||||
workgroupShared[localIndex] = initial;
|
||||
workgroupBarrier();
|
||||
|
|
@ -66,14 +69,14 @@ const createInstanceNormProgramInfo =
|
|||
workgroupBarrier();
|
||||
}
|
||||
if (localIndex == 0) {
|
||||
meanShared = workgroupShared[0] / ${dataType}(normSize);
|
||||
meanShared = ${sumVector('workgroupShared[0]', components)} / f32(normSize);
|
||||
}
|
||||
workgroupBarrier();
|
||||
|
||||
// reinitialize workgroup memory.
|
||||
initial = 0;
|
||||
for (var h = localIndex; h < normSize; h += workgroupSize) {
|
||||
let deviation = ${x.get('batch', 'channel', 'h')} - meanShared;
|
||||
initial = ${f32Type}(0);
|
||||
for (var h = localIndex; h < ${normPackedSize}; h += workgroupSize) {
|
||||
let deviation = ${f32Type}(${x.get('batch', 'channel', 'h')}) - ${f32Type}(meanShared);
|
||||
initial = initial + deviation * deviation;
|
||||
}
|
||||
workgroupShared[localIndex] = initial;
|
||||
|
|
@ -87,15 +90,16 @@ const createInstanceNormProgramInfo =
|
|||
workgroupBarrier();
|
||||
}
|
||||
if (localIndex == 0) {
|
||||
squaredNormShared = workgroupShared[0];
|
||||
squaredNormShared = ${sumVector('workgroupShared[0]', components)};
|
||||
}
|
||||
workgroupBarrier();
|
||||
|
||||
let invStdDev = 1 / sqrt(squaredNormShared / ${dataType}(normSize) + epsilon);
|
||||
let channelScale = invStdDev * ${scale.getByOffset('channel')};
|
||||
let channelShift = ${bias.getByOffset('channel')} - meanShared * channelScale;
|
||||
for (var h = localIndex; h < normSize; h += workgroupSize) {
|
||||
let value = ${x.get('batch', 'channel', 'h')} * channelScale + channelShift;
|
||||
let invStdDev = 1 / sqrt(squaredNormShared / f32(normSize) + epsilon);
|
||||
let channelScale = invStdDev * f32(${scale.getByOffset('channel')});
|
||||
let channelShift = f32(${bias.getByOffset('channel')}) - meanShared * channelScale;
|
||||
for (var h = localIndex; h < ${normPackedSize}; h += workgroupSize) {
|
||||
let value = ${x.get('batch', 'channel', 'h')} * ${dataType}(${f32Type}(channelScale)) + ${dataType}(${
|
||||
f32Type}(channelShift));
|
||||
${output.set('batch', 'channel', 'h', 'value')};
|
||||
}
|
||||
}`;
|
||||
|
|
|
|||
Loading…
Reference in a new issue