【Unity】万人同屏, 从入门到放弃之——自定义BatchRendererGroup合批渲染海量物体

avatar
作者
猴君
阅读量:0

Unity万人同屏动态避障 GPU动画 Entities Graphics高性能合批渲染插件的使用_哔哩哔哩_bilibili

  万人同屏方案(gpu动画、渲染、索敌、避障等功能),可某宝搜【店铺】:【游戏开发资源商店】获取整套方案源码。

划重点!!此方案是绕开Entities(ECS),不用写一行ECS代码,现有MonoBehavior开发工作流享受Entities渲染的性能。已有项目也能使用此方案开挂,无需代码重构!

万人同屏对抗demo测试包下载,以及万人同屏方案性能测试,有需要的老板可以某宝【搜店铺】:【游戏开发资源商店】:https://pan.baidu.com/s/1ML0DC8s0RkkTAraN9maltw?pwd=blueicon-default.png?t=N7T8https://pan.baidu.com/s/1ML0DC8s0RkkTAraN9maltw?pwd=blue

由于Dots的限制太多,对于需要dlc或热更的项目来说,Dots就爱莫能助。能不能不用Entities,只用Entities Graphics呢?

当然是可以的,Entities Graphics背后使用的接口就是Batch Renderer Group; 

自定义BatchRenderGroup合批渲染, 可以参考Unity官方文档:Initializing a BatchRendererGroup object - Unity 手册

1. 创建一个BatchRenderGroup对象和Graphics Buffer:

m_BRG = new BatchRendererGroup(this.OnPerformCulling, IntPtr.Zero);

m_InstanceData = new GraphicsBuffer(GraphicsBuffer.Target.Raw,
            BufferCountForInstances(kBytesPerInstance, kNumInstances, kExtraBytes),
            sizeof(int));

2. 注册需要渲染的Mesh和对应的Material:

m_MeshID = m_BRG.RegisterMesh(mesh);
m_MaterialID = m_BRG.RegisterMaterial(material); 

3. 为所有需要渲染的目标创建矩阵数组并传入Graphics Buffer里:

 m_InstanceData.SetData(objectToWorld, 0, (int)(byteAddressObjectToWorld / kSizeOfPackedMatrix), objectToWorld.Length);
        m_InstanceData.SetData(worldToObject, 0, (int)(byteAddressWorldToObject / kSizeOfPackedMatrix), worldToObject.Length);

4. 把Graphics Buffer添加到BatchRenderGroup进行批次渲染:

m_BatchID = m_BRG.AddBatch(metadata, m_InstanceData.bufferHandle);

创建BatchRenderGroup需要指定一个OnPerformCulling,在相机Culling时自动回调,这里可以直接使用Unity手册里的示例代码:Creating draw commands - Unity 手册

这里我主要测试的使用BatchRenderGroup合批渲染的性能,使用Job System多线程并行修改矩阵数组的位置和旋转,以控制小人移动起来。

控制小人移动的Job System代码如下:

[BurstCompile] partial struct RandomMoveJob : IJobParallelFor {     [ReadOnly]     public Unity.Mathematics.Random random;     [ReadOnly]     public float4 randomPostionRange;     [ReadOnly]     public float m_DeltaTime;      public NativeArray<Matrix4x4> matrices;     public NativeArray<float3> targetMovePoints;     public NativeArray<PackedMatrix> obj2WorldArr;     public NativeArray<PackedMatrix> world2ObjArr;     [BurstCompile]     public void Execute(int index)     {         float3 curPos = matrices[index].GetPosition();         float3 dir = targetMovePoints[index] - curPos;         if (Unity.Mathematics.math.lengthsq(dir) < 0.4f)         {             var newTargetPos = targetMovePoints[index];             newTargetPos.x = random.NextFloat(randomPostionRange.x, randomPostionRange.y);             newTargetPos.z = random.NextFloat(randomPostionRange.z, randomPostionRange.w);             targetMovePoints[index] = newTargetPos;         }          dir = math.normalizesafe(targetMovePoints[index] - curPos, Vector3.forward);         curPos += dir * m_DeltaTime;// math.lerp(curPos, targetMovePoints[index], m_DeltaTime);          var mat = matrices[index];         mat.SetTRS(curPos, Quaternion.LookRotation(dir), Vector3.one);         matrices[index] = mat;         var item = obj2WorldArr[index];         item.SetData(mat);         obj2WorldArr[index] = item;          item = world2ObjArr[index];         item.SetData(mat.inverse);         world2ObjArr[index] = item;     } }

然后在主线程Update每帧Jobs逻辑,把Jobs运算结果传入Graphics Buffer更新即可:

private void Update()     {         NativeArray<Matrix4x4> tempMatrices = new NativeArray<Matrix4x4>(matrices, Allocator.TempJob);         NativeArray<float3> tempTargetPoints = new NativeArray<float3>(m_TargetPoints, Allocator.TempJob);//worldToObject         NativeArray<PackedMatrix> tempobjectToWorldArr = new NativeArray<PackedMatrix>(matrices.Length, Allocator.TempJob);         NativeArray<PackedMatrix> tempWorldToObjectArr = new NativeArray<PackedMatrix>(matrices.Length, Allocator.TempJob);         random = new Unity.Mathematics.Random((uint)Time.frameCount);         var moveJob = new RandomMoveJob         {             matrices = tempMatrices,             targetMovePoints = tempTargetPoints,             random = random,             m_DeltaTime = Time.deltaTime * 4f,             randomPostionRange = m_randomRange,             obj2WorldArr = tempobjectToWorldArr,             world2ObjArr = tempWorldToObjectArr         };         var moveJobHandle = moveJob.Schedule(tempMatrices.Length, 64);         moveJobHandle.Complete();         matrices = moveJob.matrices.ToArray();         m_TargetPoints = moveJob.targetMovePoints.ToArray();         m_InstanceData.SetData(moveJob.obj2WorldArr, 0, (int)(byteAddressObjectToWorld / kSizeOfPackedMatrix), objectToWorld.Length);         m_InstanceData.SetData(moveJob.world2ObjArr, 0, (int)(byteAddressWorldToObject / kSizeOfPackedMatrix), worldToObject.Length);         tempMatrices.Dispose();         tempTargetPoints.Dispose();         tempobjectToWorldArr.Dispose();         tempWorldToObjectArr.Dispose();     }

Okay,跑起来看看:

 瞬间惊呆了,你没看错,使用Batch Renderer Group创建一万的小人居然能跑600多帧!!! 

难道万人同屏要成了?继续加大药量,创建10万个带有移动逻辑的小人:

 10万个奔跑的3D人物,仍然有100帧以上,有23个线程并行计算移动。

看看性能分析:

 当数量级庞大时,即使Job System + Burst编译再怎么开挂,主线程也会拖后腿的。

100万的压迫感,虽然已经成PPT了:

 难道万人同屏行业难题的门槛就这么被Unity Dots打下来了??

非也,上移动端测试:

同样1万个小人,PC端能达到惊人的600帧,而Android最强骁龙8 Gen2只有10多帧,而且工作线程数才5个; 当数量3000人时,手机端帧数46帧左右,相比传统方式没有任何提升!没错,没有任何提升。

Profiler中可以看到,耗时主要来自GPU等待CPU组织和上传渲染数据。 而Entities Graphics内部也是通过Batch Renderer Group接口实现,由此可以推断,被吹爆的Entities在移动端因该也是"水土不服":

 结论:

目前为止,我认为使用自定义BatchRendererGroup合批是PC端万人同屏的最优解了。

但是手机端性能瓶颈任重道远。手机端放弃!

最后附上本文BatchRendererGroup测试代码:

using System; using TMPro; using Unity.Burst; using Unity.Collections; using Unity.Collections.LowLevel.Unsafe; using Unity.Jobs; using Unity.Jobs.LowLevel.Unsafe; using Unity.Mathematics; using UnityEngine; using UnityEngine.Rendering;  // The PackedMatrix is a convenience type that converts matrices into // the format that Unity-provided SRP shaders expect. struct PackedMatrix {     public float c0x;     public float c0y;     public float c0z;     public float c1x;     public float c1y;     public float c1z;     public float c2x;     public float c2y;     public float c2z;     public float c3x;     public float c3y;     public float c3z;      public PackedMatrix(Matrix4x4 m)     {         c0x = m.m00;         c0y = m.m10;         c0z = m.m20;         c1x = m.m01;         c1y = m.m11;         c1z = m.m21;         c2x = m.m02;         c2y = m.m12;         c2z = m.m22;         c3x = m.m03;         c3y = m.m13;         c3z = m.m23;     }      public void SetData(Matrix4x4 m)     {         c0x = m.m00;         c0y = m.m10;         c0z = m.m20;         c1x = m.m01;         c1y = m.m11;         c1z = m.m21;         c2x = m.m02;         c2y = m.m12;         c2z = m.m22;         c3x = m.m03;         c3y = m.m13;         c3z = m.m23;     } } public class SimpleBRGExample : MonoBehaviour {     public Mesh mesh;     public Material material;     public TextMeshProUGUI text;     public TextMeshProUGUI workerCountText;     private BatchRendererGroup m_BRG;      private GraphicsBuffer m_InstanceData;     private BatchID m_BatchID;     private BatchMeshID m_MeshID;     private BatchMaterialID m_MaterialID;      // Some helper constants to make calculations more convenient.     private const int kSizeOfMatrix = sizeof(float) * 4 * 4;     private const int kSizeOfPackedMatrix = sizeof(float) * 4 * 3;     private const int kSizeOfFloat4 = sizeof(float) * 4;     private const int kBytesPerInstance = (kSizeOfPackedMatrix * 2) + kSizeOfFloat4;     private const int kExtraBytes = kSizeOfMatrix * 2;     [SerializeField] private int kNumInstances = 20000;     [SerializeField] private int m_RowCount = 200;     private Matrix4x4[] matrices;     private PackedMatrix[] objectToWorld;     private PackedMatrix[] worldToObject;     private Vector4[] colors;      private void Start()     {         m_BRG = new BatchRendererGroup(this.OnPerformCulling, IntPtr.Zero);         m_MeshID = m_BRG.RegisterMesh(mesh);         m_MaterialID = m_BRG.RegisterMaterial(material);         AllocateInstanceDateBuffer();         PopulateInstanceDataBuffer();          text.text = kNumInstances.ToString();         random = new Unity.Mathematics.Random(1);         m_TargetPoints = new float3[kNumInstances];         var offset = new Vector3(m_RowCount, 0, Mathf.CeilToInt(kNumInstances / (float)m_RowCount)) * 0.5f;         m_randomRange = new float4(-offset.x, offset.x, -offset.z, offset.z);         for (int i = 0; i < m_TargetPoints.Length; i++)         {             var newTargetPos = new float3();             newTargetPos.x = random.NextFloat(m_randomRange.x, m_randomRange.y);             newTargetPos.z = random.NextFloat(m_randomRange.z, m_randomRange.w);             m_TargetPoints[i] = newTargetPos;         }              }      float3[] m_TargetPoints;     Unity.Mathematics.Random random;     Vector4 m_randomRange;     private uint byteAddressObjectToWorld;     private uint byteAddressWorldToObject;     private uint byteAddressColor;      private void Update()     {         NativeArray<Matrix4x4> tempMatrices = new NativeArray<Matrix4x4>(matrices, Allocator.TempJob);         NativeArray<float3> tempTargetPoints = new NativeArray<float3>(m_TargetPoints, Allocator.TempJob);//worldToObject         NativeArray<PackedMatrix> tempobjectToWorldArr = new NativeArray<PackedMatrix>(matrices.Length, Allocator.TempJob);         NativeArray<PackedMatrix> tempWorldToObjectArr = new NativeArray<PackedMatrix>(matrices.Length, Allocator.TempJob);         random = new Unity.Mathematics.Random((uint)Time.frameCount);         var moveJob = new RandomMoveJob         {             matrices = tempMatrices,             targetMovePoints = tempTargetPoints,             random = random,             m_DeltaTime = Time.deltaTime * 4f,             randomPostionRange = m_randomRange,             obj2WorldArr = tempobjectToWorldArr,             world2ObjArr = tempWorldToObjectArr         };         var moveJobHandle = moveJob.Schedule(tempMatrices.Length, 64);         moveJobHandle.Complete();         matrices = moveJob.matrices.ToArray();         m_TargetPoints = moveJob.targetMovePoints.ToArray();         m_InstanceData.SetData(moveJob.obj2WorldArr, 0, (int)(byteAddressObjectToWorld / kSizeOfPackedMatrix), objectToWorld.Length);         m_InstanceData.SetData(moveJob.world2ObjArr, 0, (int)(byteAddressWorldToObject / kSizeOfPackedMatrix), worldToObject.Length);         tempMatrices.Dispose();         tempTargetPoints.Dispose();         tempobjectToWorldArr.Dispose();         tempWorldToObjectArr.Dispose();          workerCountText.text = $"JobWorkerCount:{JobsUtility.JobWorkerCount}";     }      private void AllocateInstanceDateBuffer()     {         m_InstanceData = new GraphicsBuffer(GraphicsBuffer.Target.Raw,             BufferCountForInstances(kBytesPerInstance, kNumInstances, kExtraBytes),             sizeof(int));     }     private void RefreshData()     {         m_InstanceData.SetData(objectToWorld, 0, (int)(byteAddressObjectToWorld / kSizeOfPackedMatrix), objectToWorld.Length);         m_InstanceData.SetData(worldToObject, 0, (int)(byteAddressWorldToObject / kSizeOfPackedMatrix), worldToObject.Length);     }     private void PopulateInstanceDataBuffer()     {         // Place a zero matrix at the start of the instance data buffer, so loads from address 0 return zero.         var zero = new Matrix4x4[1] { Matrix4x4.zero };          // Create transform matrices for three example instances.         matrices = new Matrix4x4[kNumInstances];         // Convert the transform matrices into the packed format that shaders expects.         objectToWorld = new PackedMatrix[kNumInstances];         // Also create packed inverse matrices.         worldToObject = new PackedMatrix[kNumInstances];         // Make all instances have unique colors.         colors = new Vector4[kNumInstances];          var offset = new Vector3(m_RowCount, 0, Mathf.CeilToInt(kNumInstances / (float)m_RowCount)) * 0.5f;         for (int i = 0; i < kNumInstances; i++)         {             matrices[i] = Matrix4x4.Translate(new Vector3(i % m_RowCount, 0, i / m_RowCount) - offset);             objectToWorld[i] = new PackedMatrix(matrices[i]);             worldToObject[i] = new PackedMatrix(matrices[0].inverse);             colors[i] = UnityEngine.Random.ColorHSV();         }          // In this simple example, the instance data is placed into the buffer like this:         // Offset | Description         //      0 | 64 bytes of zeroes, so loads from address 0 return zeroes         //     64 | 32 uninitialized bytes to make working with SetData easier, otherwise unnecessary         //     96 | unity_ObjectToWorld, three packed float3x4 matrices         //    240 | unity_WorldToObject, three packed float3x4 matrices         //    384 | _BaseColor, three float4s          // Calculates start addresses for the different instanced properties. unity_ObjectToWorld starts at          // address 96 instead of 64 which means 32 bits are left uninitialized. This is because the          // computeBufferStartIndex parameter requires the start offset to be divisible by the size of the source         // array element type. In this case, it's the size of PackedMatrix, which is 48.         byteAddressObjectToWorld = kSizeOfPackedMatrix * 2;         byteAddressWorldToObject = byteAddressObjectToWorld + kSizeOfPackedMatrix * (uint)kNumInstances;         byteAddressColor = byteAddressWorldToObject + kSizeOfPackedMatrix * (uint)kNumInstances;          // Upload the instance data to the GraphicsBuffer so the shader can load them.         m_InstanceData.SetData(zero, 0, 0, 1);         m_InstanceData.SetData(objectToWorld, 0, (int)(byteAddressObjectToWorld / kSizeOfPackedMatrix), objectToWorld.Length);         m_InstanceData.SetData(worldToObject, 0, (int)(byteAddressWorldToObject / kSizeOfPackedMatrix), worldToObject.Length);         m_InstanceData.SetData(colors, 0, (int)(byteAddressColor / kSizeOfFloat4), colors.Length);          // Set up metadata values to point to the instance data. Set the most significant bit 0x80000000 in each         // which instructs the shader that the data is an array with one value per instance, indexed by the instance index.         // Any metadata values that the shader uses and not set here will be zero. When such a value is used with         // UNITY_ACCESS_DOTS_INSTANCED_PROP (i.e. without a default), the shader interprets the         // 0x00000000 metadata value and loads from the start of the buffer. The start of the buffer which is         // is a zero matrix so this sort of load is guaranteed to return zero, which is a reasonable default value.         var metadata = new NativeArray<MetadataValue>(3, Allocator.Temp);         metadata[0] = new MetadataValue { NameID = Shader.PropertyToID("unity_ObjectToWorld"), Value = 0x80000000 | byteAddressObjectToWorld, };         metadata[1] = new MetadataValue { NameID = Shader.PropertyToID("unity_WorldToObject"), Value = 0x80000000 | byteAddressWorldToObject, };         metadata[2] = new MetadataValue { NameID = Shader.PropertyToID("_BaseColor"), Value = 0x80000000 | byteAddressColor, };          // Finally, create a batch for the instances, and make the batch use the GraphicsBuffer with the         // instance data, as well as the metadata values that specify where the properties are.          m_BatchID = m_BRG.AddBatch(metadata, m_InstanceData.bufferHandle);     }      // Raw buffers are allocated in ints. This is a utility method that calculates     // the required number of ints for the data.     int BufferCountForInstances(int bytesPerInstance, int numInstances, int extraBytes = 0)     {         // Round byte counts to int multiples         bytesPerInstance = (bytesPerInstance + sizeof(int) - 1) / sizeof(int) * sizeof(int);         extraBytes = (extraBytes + sizeof(int) - 1) / sizeof(int) * sizeof(int);         int totalBytes = bytesPerInstance * numInstances + extraBytes;         return totalBytes / sizeof(int);     }       private void OnDisable()     {         m_BRG.Dispose();     }      public unsafe JobHandle OnPerformCulling(         BatchRendererGroup rendererGroup,         BatchCullingContext cullingContext,         BatchCullingOutput cullingOutput,         IntPtr userContext)     {         // UnsafeUtility.Malloc() requires an alignment, so use the largest integer type's alignment         // which is a reasonable default.         int alignment = UnsafeUtility.AlignOf<long>();          // Acquire a pointer to the BatchCullingOutputDrawCommands struct so you can easily         // modify it directly.         var drawCommands = (BatchCullingOutputDrawCommands*)cullingOutput.drawCommands.GetUnsafePtr();         // Allocate memory for the output arrays. In a more complicated implementation, you would calculate         // the amount of memory to allocate dynamically based on what is visible.         // This example assumes that all of the instances are visible and thus allocates         // memory for each of them. The necessary allocations are as follows:         // - a single draw command (which draws kNumInstances instances)         // - a single draw range (which covers our single draw command)         // - kNumInstances visible instance indices.         // You must always allocate the arrays using Allocator.TempJob.         drawCommands->drawCommands = (BatchDrawCommand*)UnsafeUtility.Malloc(UnsafeUtility.SizeOf<BatchDrawCommand>(), alignment, Allocator.TempJob);         drawCommands->drawRanges = (BatchDrawRange*)UnsafeUtility.Malloc(UnsafeUtility.SizeOf<BatchDrawRange>(), alignment, Allocator.TempJob);         drawCommands->visibleInstances = (int*)UnsafeUtility.Malloc(kNumInstances * sizeof(int), alignment, Allocator.TempJob);         drawCommands->drawCommandPickingInstanceIDs = null;          drawCommands->drawCommandCount = 1;         drawCommands->drawRangeCount = 1;         drawCommands->visibleInstanceCount = kNumInstances;          // This example doens't use depth sorting, so it leaves instanceSortingPositions as null.         drawCommands->instanceSortingPositions = null;         drawCommands->instanceSortingPositionFloatCount = 0;          // Configure the single draw command to draw kNumInstances instances         // starting from offset 0 in the array, using the batch, material and mesh         // IDs registered in the Start() method. It doesn't set any special flags.         drawCommands->drawCommands[0].visibleOffset = 0;         drawCommands->drawCommands[0].visibleCount = (uint)kNumInstances;         drawCommands->drawCommands[0].batchID = m_BatchID;         drawCommands->drawCommands[0].materialID = m_MaterialID;         drawCommands->drawCommands[0].meshID = m_MeshID;         drawCommands->drawCommands[0].submeshIndex = 0;         drawCommands->drawCommands[0].splitVisibilityMask = 0xff;         drawCommands->drawCommands[0].flags = 0;         drawCommands->drawCommands[0].sortingPosition = 0;          // Configure the single draw range to cover the single draw command which         // is at offset 0.         drawCommands->drawRanges[0].drawCommandsBegin = 0;         drawCommands->drawRanges[0].drawCommandsCount = 1;          // This example doesn't care about shadows or motion vectors, so it leaves everything         // at the default zero values, except the renderingLayerMask which it sets to all ones         // so Unity renders the instances regardless of mask settings.         drawCommands->drawRanges[0].filterSettings = new BatchFilterSettings { renderingLayerMask = 0xffffffff, };          // Finally, write the actual visible instance indices to the array. In a more complicated         // implementation, this output would depend on what is visible, but this example         // assumes that everything is visible.         for (int i = 0; i < kNumInstances; ++i)             drawCommands->visibleInstances[i] = i;          // This simple example doesn't use jobs, so it returns an empty JobHandle.         // Performance-sensitive applications are encouraged to use Burst jobs to implement         // culling and draw command output. In this case, this function returns a         // handle here that completes when the Burst jobs finish.         return new JobHandle();     } } [BurstCompile] partial struct RandomMoveJob : IJobParallelFor {     [ReadOnly]     public Unity.Mathematics.Random random;     [ReadOnly]     public float4 randomPostionRange;     [ReadOnly]     public float m_DeltaTime;      public NativeArray<Matrix4x4> matrices;     public NativeArray<float3> targetMovePoints;     public NativeArray<PackedMatrix> obj2WorldArr;     public NativeArray<PackedMatrix> world2ObjArr;     [BurstCompile]     public void Execute(int index)     {         float3 curPos = matrices[index].GetPosition();         float3 dir = targetMovePoints[index] - curPos;         if (Unity.Mathematics.math.lengthsq(dir) < 0.4f)         {             var newTargetPos = targetMovePoints[index];             newTargetPos.x = random.NextFloat(randomPostionRange.x, randomPostionRange.y);             newTargetPos.z = random.NextFloat(randomPostionRange.z, randomPostionRange.w);             targetMovePoints[index] = newTargetPos;         }          dir = math.normalizesafe(targetMovePoints[index] - curPos, Vector3.forward);         curPos += dir * m_DeltaTime;// math.lerp(curPos, targetMovePoints[index], m_DeltaTime);          var mat = matrices[index];         mat.SetTRS(curPos, Quaternion.LookRotation(dir), Vector3.one);         matrices[index] = mat;         var item = obj2WorldArr[index];         item.SetData(mat);         obj2WorldArr[index] = item;          item = world2ObjArr[index];         item.SetData(mat.inverse);         world2ObjArr[index] = item;     } }

广告一刻

为您即时展示最新活动产品广告消息,让您随时掌握产品活动新动态!