Improve mesh buffer for avoid iOS14 performance drop (#10015)

* Improve mesh buffer for avoid iOS14 performance drop

* Fix mesh buffer unit test
This commit is contained in:
pandamicro
2022-01-18 23:13:20 +08:00
committed by GitHub
parent 59acd7eec7
commit be2b44748b
3 changed files with 106 additions and 68 deletions
-14
View File
@@ -290,13 +290,6 @@ export class Batcher2D implements IBatcher {
accessor = new StaticVBAccessor(this.device, attributes);
this._bufferAccessors.set(strideBytes, accessor);
}
// let meshBufferUseCount = this._meshBufferUseCount.get(strideBytes) || 0;
// // @ts-expect-error Property '__isWebIOS14OrIPadOS14Env' does not exist on 'sys'
// if (vertexCount && indexCount || sys.__isWebIOS14OrIPadOS14Env) {
// // useCount++ when _recreateMeshBuffer
// meshBufferUseCount++;
// }
// this._meshBufferUseCount.set(strideBytes, meshBufferUseCount);
this._staticVBBuffer = accessor;
this._currBID = -1;
@@ -535,13 +528,6 @@ export class Batcher2D implements IBatcher {
curDrawBatch.fillPasses(mat, depthStencil, dssHash, blendState, bsHash, null, this);
this._batches.push(curDrawBatch);
// HACK: After sharing buffer between drawcalls, the performance degradation a lots on iOS 14 or iPad OS 14 device
// TODO: Maybe it can be removed after Apple fixes it?
// @ts-expect-error Property '__isWebIOS14OrIPadOS14Env' does not exist on 'sys'
if (sys.__isWebIOS14OrIPadOS14Env && !this._currIsStatic) {
// this._currMeshBuffer = null;
}
}
/**
+9
View File
@@ -54,6 +54,15 @@ replaceProperty(MeshBuffer.prototype, 'MeshBuffer', [
},
]);
removeProperty(MeshBuffer.prototype, 'MeshBuffer', [
{
name: 'vertexBuffers',
},
{
name: 'indexBuffer',
},
]);
replaceProperty(Batcher2D.prototype, 'Batcher2D', [
{
name: 'currBufferBatch',
+97 -54
View File
@@ -30,15 +30,19 @@
import { Device, BufferUsageBit, MemoryUsageBit, Attribute, Buffer, BufferInfo, InputAssembler, InputAssemblerInfo } from '../../core/gfx';
import { getComponentPerVertex } from './vertex-format';
import { getError, warnID } from '../../core/platform/debug';
import { macro } from '../../core';
import { macro, sys, nextPow2 } from '../../core';
import { assertIsTrue } from '../../core/data/utils/asserts';
interface IIARef {
ia: InputAssembler;
vertexBuffers: Buffer[];
indexBuffer: Buffer;
}
export class MeshBuffer {
public static IB_SCALE = 4; // ib size scale based on vertex count
get attributes () { return this._attributes; }
get vertexFormatBytes () { return this._vertexFormatBytes; }
get vertexBuffers (): Readonly<Buffer[]> { return this._vertexBuffers; }
get indexBuffer () { return this._indexBuffer; }
public byteOffset = 0;
public vertexOffset = 0;
@@ -48,47 +52,29 @@ export class MeshBuffer {
private _dirty = false;
private _vertexFormatBytes = 0;
private _floatsPerVertex = 0;
private _initVDataCount = 0;
private _initIDataCount = 0;
private _attributes: Attribute[] = null!;
private _vertexBuffers: Buffer[] = [];
private _indexBuffer: Buffer = null!;
// InputAssembler pools for each mesh buffer, array offset correspondent
private _iaPool: InputAssembler[] = [];
private _iaPool: IIARef[] = [];
private _iaInfo: InputAssemblerInfo = null!;
private _nextFreeIAHandle = 0;
public initialize (device: Device, attrs: Attribute[]) {
const floatCount = getComponentPerVertex(attrs);
const vbStride = this._vertexFormatBytes = floatCount * Float32Array.BYTES_PER_ELEMENT;
const ibStride = Uint16Array.BYTES_PER_ELEMENT;
this._initVDataCount = macro.BATCHER2D_MEM_INCREMENT * 1024 / Float32Array.BYTES_PER_ELEMENT;
assertIsTrue(this._initVDataCount < 65536, getError(9005));
this._initIDataCount = this._initVDataCount * MeshBuffer.IB_SCALE;
this._attributes = attrs;
this._floatsPerVertex = getComponentPerVertex(attrs);
this._vertexBuffers[0] = device.createBuffer(new BufferInfo(
BufferUsageBit.VERTEX | BufferUsageBit.TRANSFER_DST,
MemoryUsageBit.HOST | MemoryUsageBit.DEVICE,
vbStride,
vbStride,
));
this._indexBuffer = device.createBuffer(new BufferInfo(
BufferUsageBit.INDEX | BufferUsageBit.TRANSFER_DST,
MemoryUsageBit.HOST | MemoryUsageBit.DEVICE,
ibStride,
ibStride,
));
// for recycle pool using purpose --
if (!this.vData || !this.iData) {
this.vData = new Float32Array(this._initVDataCount);
this.iData = new Uint16Array(this._initIDataCount);
}
// ----------
this._iaInfo = new InputAssemblerInfo(this._attributes, this._vertexBuffers, this.indexBuffer);
// Initialize the first ia
this._iaPool.push(this.createNewIA(device));
}
public reset () {
@@ -99,22 +85,20 @@ export class MeshBuffer {
public destroy () {
this.reset();
this._attributes = null!;
this._iaInfo = null!;
this.vData = null!;
this.iData = null!;
if (this._vertexBuffers[0]) {
this._vertexBuffers[0].destroy();
}
this._vertexBuffers.length = 0;
if (this._indexBuffer) {
this._indexBuffer.destroy();
}
this._indexBuffer = null!;
// Destroy InputAssemblers
for (let i = 0; i < this._iaPool.length; ++i) {
this._iaPool[i].destroy();
const iaRef = this._iaPool[i];
if (iaRef.vertexBuffers[0]) {
iaRef.vertexBuffers[0].destroy();
}
if (iaRef.indexBuffer) {
iaRef.indexBuffer.destroy();
}
iaRef.ia.destroy();
}
this._iaPool.length = 0;
}
@@ -134,19 +118,22 @@ export class MeshBuffer {
public requireFreeIA (device: Device) {
if (this._iaPool.length <= this._nextFreeIAHandle) {
this._iaPool.push(device.createInputAssembler(this._iaInfo));
this._iaPool.push(this.createNewIA(device));
}
const ia = this._iaPool[this._nextFreeIAHandle++];
const ia = this._iaPool[this._nextFreeIAHandle++].ia;
return ia;
}
public recycleIA (ia: InputAssembler) {
const pool = this._iaPool;
const id = pool.indexOf(ia);
if (id >= 0 && id < this._nextFreeIAHandle) {
// Swap to recycle the ia
pool[id] = pool[--this._nextFreeIAHandle];
pool[this._nextFreeIAHandle] = ia;
for (let i = 0; i < this._nextFreeIAHandle; ++i) {
if (ia === pool[i].ia) {
// Swap to recycle the ia
const iaRef = pool[i];
pool[i] = pool[--this._nextFreeIAHandle];
pool[this._nextFreeIAHandle] = iaRef;
return;
}
}
}
@@ -164,19 +151,75 @@ export class MeshBuffer {
return;
}
const verticesData = new Float32Array(this.vData.buffer, 0, this.byteOffset >> 2);
const indicesData = new Uint16Array(this.iData.buffer, 0, this.indexOffset);
// On iOS14, different IAs can not share same GPU buffer, so must submit the same date to different buffers
// @ts-expect-error Property '__isWebIOS14OrIPadOS14Env' does not exist on 'sys'
const iOS14 = sys.__isWebIOS14OrIPadOS14Env;
const submitCount = iOS14 ? this._nextFreeIAHandle : 1;
const byteCount = this.byteOffset;
const indexCount = this.indexOffset;
for (let i = 0; i < submitCount; ++i) {
const iaRef = this._iaPool[i];
// if (iOS14) {
// indexCount = iaRef.ia.firstIndex + iaRef.ia.indexCount;
// const maxVertex = this.iData[indexCount];
// // Only upload as much data as needed, to avoid frequent resize, using pow2 size
// // Wrong implementation because maxVertex might be larger than the last vertex id, hard to find the correct max vertex
// byteCount = Math.min(this.byteOffset, nextPow2(maxVertex + 2) * this.vertexFormatBytes);
// }
const vertexBuffer = this._vertexBuffers[0];
if (this.byteOffset > vertexBuffer.size) {
vertexBuffer.resize(this.byteOffset);
}
vertexBuffer.update(verticesData);
const verticesData = new Float32Array(this.vData.buffer, 0, byteCount >> 2);
const indicesData = new Uint16Array(this.iData.buffer, 0, indexCount);
if (this.indexOffset * 2 > this.indexBuffer.size) {
this.indexBuffer.resize(this.indexOffset * 2);
const vertexBuffer = iaRef.vertexBuffers[0];
if (byteCount > vertexBuffer.size) {
vertexBuffer.resize(byteCount);
}
vertexBuffer.update(verticesData);
if (indexCount * 2 > iaRef.indexBuffer.size) {
iaRef.indexBuffer.resize(indexCount * 2);
}
iaRef.indexBuffer.update(indicesData);
}
this._indexBuffer.update(indicesData);
this._dirty = false;
}
private createNewIA (device: Device): IIARef {
let ia;
let vertexBuffers;
let indexBuffer;
// HACK: After sharing buffer between drawcalls, the performance degradation a lots on iOS 14 or iPad OS 14 device
// TODO: Maybe it can be removed after Apple fixes it?
// @ts-expect-error Property '__isWebIOS14OrIPadOS14Env' does not exist on 'sys'
if (sys.__isWebIOS14OrIPadOS14Env || !this._iaPool[0]) {
const vbStride = this._vertexFormatBytes = this._floatsPerVertex * Float32Array.BYTES_PER_ELEMENT;
const ibStride = Uint16Array.BYTES_PER_ELEMENT;
const vertexBuffer = device.createBuffer(new BufferInfo(
BufferUsageBit.VERTEX | BufferUsageBit.TRANSFER_DST,
MemoryUsageBit.HOST | MemoryUsageBit.DEVICE,
vbStride,
vbStride,
));
indexBuffer = device.createBuffer(new BufferInfo(
BufferUsageBit.INDEX | BufferUsageBit.TRANSFER_DST,
MemoryUsageBit.HOST | MemoryUsageBit.DEVICE,
ibStride,
ibStride,
));
vertexBuffers = [vertexBuffer];
// Reuse purpose for new IAs
this._iaInfo = new InputAssemblerInfo(this._attributes, vertexBuffers, indexBuffer);
ia = device.createInputAssembler(this._iaInfo);
} else {
ia = device.createInputAssembler(this._iaInfo);
vertexBuffers = this._iaInfo.vertexBuffers;
indexBuffer = this._iaInfo.indexBuffer;
}
return {
ia,
vertexBuffers,
indexBuffer,
};
}
}