birkeh/o3de
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
Files
1044dc3da1e70eab8f8634d972a8dc6b072cf565
o3de/Gems/SVOGI/Code/Source/SvoTree.cpp
T

1322 lines
48 KiB
C++
Raw Blame History

/*
* All or portions of this file Copyright (c) Amazon.com, Inc. or its affiliates or
* its licensors.
*
* For complete copyright and license terms please see the LICENSE at the root of this
* distribution (the "License"). All use of this software is governed by the License,
* or, if provided, by the license below or the license accompanying this file. Do not
* remove or modify any license notices. This file is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
*
*/
// Original file Copyright Crytek GMBH or its affiliates, used under license.
// Description : CPU side SVO
#include "SVOGI_precompiled.h"
#include <AzCore/std/parallel/thread.h>
#include <AzCore/std/parallel/atomic.h>
#include <AzCore/std/parallel/lock.h>
#include "SvoTree.h"
#include "TextureBlockPacker.h"
#include "IRenderAuxGeom.h"
#include "FrameProfiler.h"
#include "MathConversion.h"
#include <AzFramework/Terrain/TerrainDataRequestBus.h>
namespace SVOGI
{
//Since the Job System has no priority or size scheduling hints we don't want to starve out other jobs with
//the longer running GI jobs. This will keep us to a 'reasonable' number of jobs.
static AZ::u32 s_svoMaxJobCount = (AZStd::thread::hardware_concurrency() / 4 > 0) ? AZStd::thread::hardware_concurrency() / 4 : 1;
//Scratch working buffer for GI calculations.
static DataBrick<GISubVoxels>* s_scratchData;
static AZStd::atomic_bool* s_freeScratch;
////////////////////////////////////////////////////////////////////////////////////////////
// SVO ENVIRONMENT
////////////////////////////////////////////////////////////////////////////////////////////
const AZ::s32 SvoEnvironment::s_uninitializedTexturePoolId = 0;
AZ::s32 SvoEnvironment::m_currentPassFrameId = 0;
AZ::s32 SvoEnvironment::m_brickTexturePoolDimXY = 128;
AZ::s32 SvoEnvironment::m_brickTexturePoolDimZ = 256;
SvoEnvironment::SvoEnvironment()
{
m_brickTextureFormat = eTF_R8G8B8A8; // eTF_BC3
AllocateTexturePools();
m_prevCheckVal = -1000000;
m_svoRoot.reset();
m_globalSpecularCM = nullptr;
m_globalSpecularCM_Mult = 1;
m_activeVoxels = 0;
m_evictGpu = false;
m_blockPacker = aznew TextureBlockPacker3D(nAtlasDimMaxXY, nAtlasDimMaxXY, nAtlasDimMaxZ, true);
m_blockIndex = 0;
s_scratchData = (DataBrick<GISubVoxels>*)azmalloc(sizeof(DataBrick<GISubVoxels>)*s_svoMaxJobCount);
s_scratchData = new(s_scratchData) DataBrick<GISubVoxels>[s_svoMaxJobCount];
s_freeScratch = new AZStd::atomic_bool[s_svoMaxJobCount];
for (AZ::s32 i = 0; i < s_svoMaxJobCount; ++i)
{
s_freeScratch[i].store(true);
}
EBUS_EVENT_RESULT(m_jobContext, AZ::JobManagerBus, GetGlobalContext);
}
SvoEnvironment::~SvoEnvironment()
{
m_voxelJobsCompletion.StartAndWaitForCompletion();
// To avoid potential memory issues during tear down, ensure that all voxels are destroyed
// before destroying the block packer as they refer to memory owned by the block packer.
m_processingQueue.clear();
m_brickUpdateQueue.clear();
m_svoRoot.reset();
delete m_blockPacker;
delete[] s_freeScratch;
azfree(s_scratchData);
SvoEnvironment::m_currentPassFrameId = 0;
DeallocateTexturePools();
}
void SvoEnvironment::ReconstructTree()
{
AZ_PROFILE_FUNCTION(AZ::Debug::ProfileCategory::Renderer);
if (gEnv->pConsole->GetCVar("e_svoTI_Active")->GetIVal())
{
m_brickUpdateQueue.clear();
m_processingQueue.clear();
m_svoRoot.reset();
DeallocateTexturePools();
AllocateTexturePools();
// Keep pending insertions and include inserted meshes to it.
// They will be reconsidered on the first update after the recreation of the root.
m_globalInsertions.insert(m_globalInsertedMeshes.begin(), m_globalInsertedMeshes.end());
m_globalInsertedMeshes.clear();
m_globalRemovals.clear();
AZ::Aabb terrainAabb = AZ::Aabb::CreateFromPoint(AZ::Vector3::CreateZero());
AzFramework::Terrain::TerrainDataRequestBus::BroadcastResult(terrainAabb, &AzFramework::Terrain::TerrainDataRequests::GetTerrainAabb);
m_svoRoot = AZStd::make_shared<Voxel>(terrainAabb, nullptr, this, 0);
}
}
//Push all nodes that need to be updated into the update queue.
void SvoEnvironment::UpdateVoxels()
{
AZ_PROFILE_FUNCTION(AZ::Debug::ProfileCategory::Renderer);
if (!m_svoRoot)
{
if (!gEnv->p3DEngine->LevelLoadingInProgress() || gEnv->IsEditor())
{
AZ::Aabb terrainAabb = AZ::Aabb::CreateFromPoint(AZ::Vector3::CreateZero());
AzFramework::Terrain::TerrainDataRequestBus::BroadcastResult(terrainAabb, &AzFramework::Terrain::TerrainDataRequests::GetTerrainAabb);
m_svoRoot = AZStd::make_shared<Voxel>(terrainAabb, nullptr, this, 0);
}
}
if (m_svoRoot)
{
float maxSize = gEnv->pConsole->GetCVar("e_svoMaxNodeSize")->GetFVal();
float minSize = gEnv->pConsole->GetCVar("e_svoMinNodeSize")->GetFVal();
bool hasNewInsertionsOrRemovals = !(m_globalInsertions.empty() && m_globalRemovals.empty());
if (hasNewInsertionsOrRemovals)
{
m_svoRoot->EnqueueMeshes(m_globalInsertions, m_globalRemovals, maxSize);
m_globalInsertedMeshes.insert(m_globalInsertions.begin(), m_globalInsertions.end());
m_globalInsertions.clear();
m_globalRemovals.clear();
}
m_svoRoot->Update(m_processingQueue, m_svoRoot, maxSize, minSize);
}
}
//Process nodes that need updating and remove outdated nodes.
void SvoEnvironment::ProcessVoxels()
{
AZ_PROFILE_FUNCTION(AZ::Debug::ProfileCategory::Renderer);
AZ::s32 maxLoadedNodes = gEnv->pConsole->GetCVar("e_svoMaxBricksOnCPU")->GetIVal();
AZ::s32 maxNodesPerJob = gEnv->pConsole->GetCVar("e_svoMaxVoxelUpdatesPerJob")->GetIVal();
float maxSize = gEnv->pConsole->GetCVar("e_svoMaxNodeSize")->GetFVal();
while (!m_processingQueue.empty())
{
//Compute Free Scratch Space location
AZ::u32 offset = 0;
for (; offset < s_svoMaxJobCount; ++offset)
{
if (s_freeScratch[offset])
{
s_freeScratch[offset].store(false);
break;
}
}
//No free scratch space stop processing this frame.
if (offset == s_svoMaxJobCount)
{
return;
}
//Build working voxel set for job
AZStd::vector<AZStd::shared_ptr<Voxel>> voxels;
voxels.reserve(maxNodesPerJob);
AZ::u32 count = 0;
while (!m_processingQueue.empty() && count < maxNodesPerJob)
{
voxels.push_back(m_processingQueue.front());
m_processingQueue.pop_front();
++count;
}
auto voxelJobFunc = [this, offset, voxels, maxSize, maxLoadedNodes]()
{
AZ_PROFILE_SCOPE(AZ::Debug::ProfileCategory::Renderer, "SvoEnvironment::ProcessVoxels:JobFunc");
for (auto voxel : voxels)
{
voxel->UpdateBrickData(maxSize, maxLoadedNodes, s_scratchData[offset]);
voxel->m_cpuEnqueued.store(false);
}
s_freeScratch[offset] = true;
};
AZ::Job* job = AZ::CreateJobFunction(voxelJobFunc, true, m_jobContext);
job->SetDependent(&m_voxelJobsCompletion);
job->Start();
}
}
void SvoEnvironment::EvictVoxels()
{
AZ_PROFILE_FUNCTION(AZ::Debug::ProfileCategory::Renderer);
if (!m_svoRoot)
{
return;
}
AZ::s32 maxLoadedNodes = gEnv->pConsole->GetCVar("e_svoMaxBricksOnCPU")->GetIVal();
if ((m_lastEvictionFrame + m_delayToEvictInFrames) < GetCurrPassMainFrameID() ||
m_activeVoxels >= maxLoadedNodes)
{
m_lastEvictionFrame = GetCurrPassMainFrameID();
float minSize = gEnv->pConsole->GetCVar("e_svoMinNodeSize")->GetFVal();
// Evict voxels older than m_evictionDelay
m_svoRoot->Evict(m_delayToEvictInFrames, false, minSize);
// Warn the user that the value of e_svoMaxBricksOnCPU is not high enough.
// NOTE: Because Evict() does not release nodes that are or will be processed, we will only
// warn the user when it keeps happening after a reasonable amount of continuous frames.
if (m_activeVoxels >= maxLoadedNodes)
{
++m_numFramesReachingMaxBricksOnCPU;
if (m_numFramesReachingMaxBricksOnCPU >= m_delayToWarnReachingMaxBricksOnCPUInFrames)
{
AZ_Warning("SVOGI", false, "Maximum number of active voxels reached (%d). Increment the value of e_svoMaxBricksOnCPU.", maxLoadedNodes);
m_numFramesReachingMaxBricksOnCPU = 0; // Avoid warn every frame
}
}
else
{
m_numFramesReachingMaxBricksOnCPU = 0;
}
}
}
void SvoEnvironment::EvictGpuData()
{
AZ_PROFILE_FUNCTION(AZ::Debug::ProfileCategory::Renderer);
if (m_evictGpu)
{
if (m_svoRoot)
{
m_svoRoot->EvictGpuData(m_evictionDelayInFrames, false);
}
m_evictGpu = false;
}
}
void SvoEnvironment::UploadVoxels(bool showVoxels)
{
AZ_PROFILE_FUNCTION(AZ::Debug::ProfileCategory::Renderer);
AZ::s32 maxBricksPerFrame = gEnv->pConsole->GetCVar("e_svoMaxBrickUpdates")->GetIVal();
AZ::s32 bricksUploaded = 0;
{
while (!m_brickUpdateQueue.empty() && bricksUploaded <= maxBricksPerFrame)
{
AZStd::shared_ptr<Voxel> voxel = m_brickUpdateQueue.front();
voxel->UpdateBrickRenderData();
voxel->UpdateTreeRenderData();
//Mark block as needing to be processed.
if (voxel->m_block)
{
voxel->m_block->m_staticDirty = true;
voxel->m_block->m_dynamicDirty = true;
}
voxel->m_gpuEnqueued = false;
m_brickUpdateQueue.pop_front();
++bricksUploaded;
}
}
if (m_svoRoot)
{
float maxSize = gEnv->pConsole->GetCVar("e_svoMaxNodeSize")->GetFVal();
m_svoRoot->ReserveGPUMemory(m_svoRoot, maxSize);
m_svoRoot->UpdateGpuTree(m_svoRoot);
if (showVoxels)
{
m_svoRoot->DrawVoxels();
}
}
}
void SvoEnvironment::GetSvoStaticTextures(I3DEngine::SSvoStaticTexInfo& svoInfo, PodArray<I3DEngine::SLightTI>* pLightsTI_S, PodArray<I3DEngine::SLightTI>* pLightsTI_D)
{
AZ_PROFILE_FUNCTION(AZ::Debug::ProfileCategory::Renderer);
svoInfo.pTexTree = gEnv->pRenderer->EF_GetTextureByID(m_nTexNodePoolId);
svoInfo.pTexOpac = gEnv->pRenderer->EF_GetTextureByID(m_nTexOpasPoolId);
svoInfo.pTexRgb0 = gEnv->pRenderer->EF_GetTextureByID(m_nTexRgb0PoolId);
svoInfo.pTexRgb1 = gEnv->pRenderer->EF_GetTextureByID(m_nTexRgb1PoolId);
svoInfo.pTexDynl = gEnv->pRenderer->EF_GetTextureByID(m_nTexDynlPoolId);
svoInfo.pTexRgb2 = gEnv->pRenderer->EF_GetTextureByID(m_nTexRgb2PoolId);
svoInfo.pTexRgb3 = gEnv->pRenderer->EF_GetTextureByID(m_nTexRgb3PoolId);
svoInfo.pTexNorm = gEnv->pRenderer->EF_GetTextureByID(m_nTexNormPoolId);
svoInfo.pTexAldi = gEnv->pRenderer->EF_GetTextureByID(m_nTexAldiPoolId);
svoInfo.pGlobalSpecCM = m_globalSpecularCM;
svoInfo.fGlobalSpecCM_Mult = m_globalSpecularCM_Mult;
svoInfo.nTexDimXY = m_brickTexturePoolDimXY;
svoInfo.nTexDimZ = m_brickTexturePoolDimZ;
svoInfo.nBrickSize = brickDimension;
svoInfo.bSvoReady = true;
*pLightsTI_S = m_lightsTI_S;
*pLightsTI_D = m_lightsTI_D;
}
void SvoEnvironment::GetSvoBricksForUpdate(PodArray<I3DEngine::SSvoNodeInfo>& arrNodeInfo, bool getDynamic)
{
AZ_PROFILE_FUNCTION(AZ::Debug::ProfileCategory::Renderer);
arrNodeInfo.Clear();
if (!m_blockPacker)
{
return;
}
if (!gEnv->pConsole->GetCVar("e_svoTI_Active")->GetIVal())
{
return;
}
const AZ::u32 numberOfBlocks = m_blockPacker->GetNumBlocks();
AZ::u32 maxUpdatesPerFrame = gEnv->pConsole->GetCVar("e_svoMaxBrickUpdates")->GetIVal();
AZ::u32 blocksAdded = 0;
AZ::u32 oldStart = m_blockIndex;
for (; blocksAdded < maxUpdatesPerFrame;)
{
if (TextureBlock3D* block = m_blockPacker->GetBlockInfo(m_blockIndex))
{
if ((!getDynamic && block->m_staticDirty) || (getDynamic && block->m_dynamicDirty))
{
I3DEngine::SSvoNodeInfo nodeInfo;
nodeInfo.wsBox = AZAabbToLyAABB(block->m_worldBox);
nodeInfo.tcBox = AZAabbToLyAABB(block->m_textureBox);
nodeInfo.nAtlasOffset = block->m_atlasOffset;
++blocksAdded;
arrNodeInfo.Add(nodeInfo);
if (!getDynamic)
{
block->m_staticDirty = false;
}
else
{
block->m_dynamicDirty = false;
}
block->m_lastUpdatedFrame = GetCurrPassMainFrameID();
}
}
m_blockIndex = (m_blockIndex + 1) % numberOfBlocks;
//Walked full block list back to where we started. Break for now.
if (m_blockIndex == oldStart)
{
break;
}
}
//Due to the fact that we are not tracking lighting changes
//refresh the block if it hasn't been updated in a while.
//Once we have a system for detecting if a light has changed
//with respect to a given voxel region we can remove this behavior.
for (AZ::u32 blockIndex = 0; blockIndex < numberOfBlocks; blockIndex++)
{
if (TextureBlock3D* block = m_blockPacker->GetBlockInfo(blockIndex))
{
if (GetCurrPassMainFrameID() - block->m_lastUpdatedFrame > m_blockUpdateDelayInFrames)
{
block->m_staticDirty = true;
block->m_dynamicDirty = true;
}
}
}
}
static AZ::s32 SLightTI_Compare(const void* v1, const void* v2)
{
I3DEngine::SLightTI* p[2] = { (I3DEngine::SLightTI*)v1, (I3DEngine::SLightTI*)v2 };
if (p[0]->fSortVal > p[1]->fSortVal)
{
return 1;
}
if (p[0]->fSortVal < p[1]->fSortVal)
{
return -1;
}
return 0;
}
void SvoEnvironment::CollectLights()
{
AZ_PROFILE_FUNCTION(AZ::Debug::ProfileCategory::Renderer);
AABB nodeBox;
nodeBox.Reset();
nodeBox.Add(gEnv->pSystem->GetViewCamera().GetPosition());
nodeBox.Expand(Vec3(256, 256, 256));
m_lightsTI_S.Clear();
m_lightsTI_D.Clear();
if (AZ::s32 nCount = gEnv->p3DEngine->GetObjectsByTypeInBox(eERType_Light, nodeBox, (IRenderNode**)0))
{
AZStd::vector<IRenderNode*> arrObjects(nCount);
nCount = gEnv->p3DEngine->GetObjectsByTypeInBox(eERType_Light, nodeBox, &arrObjects[0]);
for (AZ::s32 nL = 0; nL < nCount; nL++)
{
ILightSource* pRN = (ILightSource*)arrObjects[nL];
CDLight& rLight = pRN->GetLightProperties();
I3DEngine::SLightTI lightTI;
memset(&lightTI, 0, sizeof(lightTI));
IRenderNode::EVoxelGIMode eVoxMode = pRN->GetVoxelGIMode();
if (eVoxMode)
{
lightTI.vPosR = Vec4(rLight.m_Origin, rLight.m_fRadius);
if ((rLight.m_Flags & DLF_PROJECT) && (rLight.m_fLightFrustumAngle < 90.f) && rLight.m_pLightImage)
{
lightTI.vDirF = Vec4(pRN->GetMatrix().GetColumn(0), rLight.m_fLightFrustumAngle * 2);
}
else
{
lightTI.vDirF = Vec4(0, 0, 0, 0);
}
if (eVoxMode == IRenderNode::VM_Dynamic)
{
lightTI.vCol = rLight.m_Color.toVec4();
}
else
{
lightTI.vCol = rLight.m_BaseColor.toVec4();
}
lightTI.vCol.w = (rLight.m_Flags & DLF_CASTSHADOW_MAPS) ? 1.f : 0.f;
if (rLight.m_Flags & DLF_SUN)
{
lightTI.fSortVal = -1;
}
else
{
Vec3 vCamPos = m_camera.GetPosition();
lightTI.fSortVal = vCamPos.GetDistance(rLight.m_Origin) / max(24.f, rLight.m_fRadius);
}
if (eVoxMode == IRenderNode::VM_Dynamic)
{
if ((pRN->GetDrawFrame(0) > 10) && (pRN->GetDrawFrame(0) >= (AZ::s32)GetCurrPassMainFrameID()))
{
m_lightsTI_D.Add(lightTI);
}
}
else
{
m_lightsTI_S.Add(lightTI);
}
}
}
if (m_lightsTI_S.Count() > 1)
{
qsort(m_lightsTI_S.GetElements(), m_lightsTI_S.Count(), sizeof(m_lightsTI_S[0]), SLightTI_Compare);
}
if (m_lightsTI_D.Count() > 1)
{
qsort(m_lightsTI_D.GetElements(), m_lightsTI_D.Count(), sizeof(m_lightsTI_D[0]), SLightTI_Compare);
}
if (m_lightsTI_D.Count() > 8)
{
m_lightsTI_D.PreAllocate(8);
}
}
m_globalSpecularCM = 0;
m_globalSpecularCM_Mult = 0;
AZ::Aabb terrainAabb = AZ::Aabb::CreateFromPoint(AZ::Vector3::CreateZero());
AzFramework::Terrain::TerrainDataRequestBus::BroadcastResult(terrainAabb, &AzFramework::Terrain::TerrainDataRequests::GetTerrainAabb);
AABB areaBox = AZAabbToLyAABB(terrainAabb);
if (AZ::s32 nCount = gEnv->p3DEngine->GetObjectsByTypeInBox(eERType_Light, areaBox, (IRenderNode**)0))
{
AZStd::vector<IRenderNode*> arrObjects(nCount, nullptr);
nCount = gEnv->p3DEngine->GetObjectsByTypeInBox(eERType_Light, areaBox, &arrObjects[0]);
float fMaxRadius = 999;
for (AZ::s32 nL = 0; nL < nCount; nL++)
{
ILightSource* pRN = (ILightSource*)arrObjects[nL];
CDLight& rLight = pRN->GetLightProperties();
if (rLight.m_fRadius > fMaxRadius && rLight.m_Flags & DLF_DEFERRED_CUBEMAPS)
{
fMaxRadius = rLight.m_fRadius;
m_globalSpecularCM = rLight.GetSpecularCubemap();
m_globalSpecularCM_Mult = rLight.m_SpecMult;
}
}
}
}
void SvoEnvironment::UpsertMesh(AZ::EntityId entityId, AZ::Transform transform, AZ::Aabb worldAabb,
AZ::Data::Asset<LmbrCentral::MeshAsset> meshAsset, _smart_ptr<IMaterial> material)
{
// Both cases insertion and update of a mesh are treated the same
// way, as a new insertion, so remove the mesh first.
RemoveMesh(entityId);
AZStd::shared_ptr<MeshData> data = AZStd::make_shared<MeshData>(entityId, transform, worldAabb, meshAsset, material);
[[maybe_unused]] bool inserted = m_globalInsertions.insert({ entityId, data }).second;
AZ_Assert(inserted, "Tried to double insert mesh.");
// NOTE: At this point the mesh will be added to:
// - m_globalInsertions
// - m_globalRemovals (if the mesh was already inserted, this is a mesh update)
}
void SvoEnvironment::RemoveMesh(AZ::EntityId entityId)
{
//Check if mesh is in the list of current objects
auto searchInserted = m_globalInsertedMeshes.find(entityId);
if (searchInserted != m_globalInsertedMeshes.end())
{
m_globalRemovals.insert(*searchInserted);
m_globalInsertedMeshes.erase(searchInserted);
}
//Remove pending insertions.
auto searchInsertions = m_globalInsertions.find(entityId);
if (searchInsertions != m_globalInsertions.end())
{
m_globalInsertions.erase(searchInsertions);
}
}
void SvoEnvironment::AllocateTexturePool(AZ::s32& texPoolId, AZ::s32 width, AZ::s32 height, AZ::s32 depth, ETEX_Format texFormat, AZ::s32 filter, AZ::s32 flags)
{
if (texPoolId == s_uninitializedTexturePoolId)
{
texPoolId = gEnv->pRenderer->DownLoadToVideoMemory3D(NULL, width, height, depth, texFormat, texFormat, 1, false, filter, 0, 0, flags);
}
}
void SvoEnvironment::AllocateTexturePools()
{
AZ_PROFILE_FUNCTION(AZ::Debug::ProfileCategory::Renderer);
AZ::s32 nFlagsReadOnly = FT_DONT_STREAM;
AZ::s32 nFlagsReadWrite = FT_DONT_STREAM | FT_USAGE_UNORDERED_ACCESS | FT_USAGE_UAV_RWTEXTURE;
AllocateTexturePool(m_nTexRgb0PoolId, m_brickTexturePoolDimXY, m_brickTexturePoolDimXY, m_brickTexturePoolDimZ, m_brickTextureFormat, FILTER_LINEAR, nFlagsReadWrite);
if (gEnv->pConsole->GetCVar("e_svoTI_Active")->GetIVal() &&
gEnv->pConsole->GetCVar("e_svoTI_IntegrationMode")->GetIVal())
{
// direct lighting
AllocateTexturePool(m_nTexRgb1PoolId, m_brickTexturePoolDimXY, m_brickTexturePoolDimXY, m_brickTexturePoolDimZ, m_brickTextureFormat, FILTER_LINEAR, nFlagsReadWrite);
// dyn direct lighting
AllocateTexturePool(m_nTexDynlPoolId, m_brickTexturePoolDimXY, m_brickTexturePoolDimXY, m_brickTexturePoolDimZ, m_brickTextureFormat, FILTER_LINEAR, nFlagsReadWrite);
// propagation
AllocateTexturePool(m_nTexRgb2PoolId, m_brickTexturePoolDimXY, m_brickTexturePoolDimXY, m_brickTexturePoolDimZ, m_brickTextureFormat, FILTER_LINEAR, nFlagsReadWrite);
// propagation
AllocateTexturePool(m_nTexRgb3PoolId, m_brickTexturePoolDimXY, m_brickTexturePoolDimXY, m_brickTexturePoolDimZ, m_brickTextureFormat, FILTER_LINEAR, nFlagsReadWrite);
}
AllocateTexturePool(m_nTexNormPoolId, m_brickTexturePoolDimXY, m_brickTexturePoolDimXY, m_brickTexturePoolDimZ, m_brickTextureFormat, FILTER_LINEAR, nFlagsReadWrite);
AllocateTexturePool(m_nTexAldiPoolId, m_brickTexturePoolDimXY, m_brickTexturePoolDimXY, m_brickTexturePoolDimZ, m_brickTextureFormat, FILTER_LINEAR, nFlagsReadWrite);
AllocateTexturePool(m_nTexOpasPoolId, m_brickTexturePoolDimXY, m_brickTexturePoolDimXY, m_brickTexturePoolDimZ, m_brickTextureFormat, FILTER_LINEAR, nFlagsReadWrite);
AllocateTexturePool(m_nTexNodePoolId, nVoxNodPoolDimXY, nVoxNodPoolDimXY, nVoxNodPoolDimZ, eTF_R32G32B32A32F, FILTER_POINT, nFlagsReadOnly);
}
void SvoEnvironment::DeallocateTexturePool(AZ::s32& texPoolId)
{
if (texPoolId != s_uninitializedTexturePoolId)
{
gEnv->pRenderer->RemoveTexture(texPoolId);
texPoolId = s_uninitializedTexturePoolId;
}
}
void SvoEnvironment::DeallocateTexturePools()
{
DeallocateTexturePool(m_nTexRgb0PoolId);
DeallocateTexturePool(m_nTexRgb1PoolId);
DeallocateTexturePool(m_nTexDynlPoolId);
DeallocateTexturePool(m_nTexRgb2PoolId);
DeallocateTexturePool(m_nTexRgb3PoolId);
DeallocateTexturePool(m_nTexNormPoolId);
DeallocateTexturePool(m_nTexAldiPoolId);
DeallocateTexturePool(m_nTexOpasPoolId);
DeallocateTexturePool(m_nTexNodePoolId);
}
void SvoEnvironment::SetCamera(const CCamera& newCam)
{
m_camera = newCam;
}
////////////////////////////////////////////////////////////////////////////////////////////
// Voxels
////////////////////////////////////////////////////////////////////////////////////////////
Voxel::Voxel(const AZ::Aabb& box, AZStd::shared_ptr<Voxel> parent, SvoEnvironment* env, AZ::u8 childIndex)
{
AZ_Assert(env, "Invalid SVO Environment");
AZ_Assert(childIndex < s_numVoxelChildren, "Invalid child index: %d", childIndex);
m_parentNode = parent;
m_nodeBox = box;
m_boxSize = box.GetZExtent();
m_svoEnv = env;
m_cpuEnqueued.store(false);
m_gpuEnqueued = false;
m_childIndex = childIndex;
m_childOffsetsDirty = true;
m_blockID = TextureBlockPacker3D::s_invalidBlockID;
m_block = nullptr;
m_brick = nullptr;
// Use current frame to avoid store frame 0 when voxel is created.
m_lastVisibleFrameId = GetCurrPassMainFrameID();
m_lastUpdatedFrameId = m_lastVisibleFrameId;
m_queueId = 0;
}
Voxel::~Voxel()
{
ReleaseBlock();
if (m_brick)
{
if (m_brick->HasBrickData())
{
--m_svoEnv->m_activeVoxels;
}
delete m_brick;
m_brick = nullptr;
}
}
void Voxel::ReleaseBlock()
{
if (m_blockID != TextureBlockPacker3D::s_invalidBlockID)
{
m_svoEnv->m_blockPacker->RemoveBlock(m_blockID);
m_blockID = TextureBlockPacker3D::s_invalidBlockID;
}
m_block = nullptr;
}
void Voxel::Update(AZStd::deque<AZStd::shared_ptr<Voxel>>& processingQueue, AZStd::shared_ptr<Voxel> self, float maxSize, float minSize)
{
//If the node is not visible stop updating.
if (!m_svoEnv->m_camera.IsAABBVisible_E(AZAabbToLyAABB(m_nodeBox)))
{
return;
}
m_lastVisibleFrameId = GetCurrPassMainFrameID();
//If voxel is "small" relative to camera distance do not upload data to GPU.
{
AZStd::shared_ptr<Voxel> parent = m_parentNode.lock();
bool voxelIsSmall = GetLodRatio() > m_svoEnv->m_voxelLodCutoff;
bool parentHasObjectData = parent && parent->m_brick && parent->m_brick->HasBrickData() && !parent->m_brick->m_terrainOnly;
if (voxelIsSmall && !parentHasObjectData)
{
return;
}
}
m_lastUpdatedFrameId = m_lastVisibleFrameId;
//Enqueue the node for processing if necessary.
{
//Lock to prevent buffer swapping during write
AZStd::lock_guard<AZStd::mutex> queueLock(m_queueMutex);
AZ::u32 queueId = m_queueId;
//If the voxel is not equeued and has either never been processed before (!(m_brick || m_block)) or has entities
//to be processed enqueue it.
const bool voxelNotQueued = !m_cpuEnqueued;
const bool voxelNeverProcessed = !(m_brick || m_block);
const bool hasInsertionsOrRemovals = (!m_insertions[queueId].empty() || !m_removals[queueId].empty());
if (voxelNotQueued && (hasInsertionsOrRemovals || voxelNeverProcessed))
{
processingQueue.push_back(self);
m_cpuEnqueued.store(true);
}
}
//If node has brick data generate children.
//If the node is bigger than max size then generate the children automatically.
if ((m_brick && m_brick->HasBrickData()) || m_boxSize > maxSize)
{
AllocateChildren(self, maxSize, minSize);
}
for (AZ::u32 childIndex = 0; childIndex < s_numVoxelChildren; childIndex++)
{
AZStd::shared_ptr<Voxel> child = m_children[childIndex];
if (child)
{
child->Update(processingQueue, child, maxSize, minSize);
}
}
}
void Voxel::Evict(AZ::u32 frameDelay, bool forceEvict, float minSize)
{
//This function assumes the root is not evictable.
//To evict the entire tree reset the root pointer.
AZStd::shared_ptr<Voxel> parent = m_parentNode.lock();
if (parent && (m_lastVisibleFrameId < (GetCurrPassMainFrameID() - frameDelay) || forceEvict))
{
// Remove voxel from the tree.
//
// When the last shared pointer is destroyed then this voxel will be destroyed and
// therefore all its children as well. That includes all bricks' data and gpu blocks
// of this voxel and all its children.
// The last shared pointer can be:
// - This function's caller.
// - If this voxel will be processed a shared pointer is queued in m_processingQueue or m_brickUpdateQueue.
// - If this Voxel is being processed by a job a shared pointer will be in vector "voxels" (see function SvoEnvironment::ProcessVoxels()).
//
// NOTE: It's possible to remove the voxel from m_processingQueue and m_brickUpdateQueue now
// to get the memory back faster, but it would not be accurate unless walking its children to
// remove them as well. The cost of walking the children and remove them from the queues is not
// worth it, eventually they will be processed and destroyed in a few frames anyway.
parent->m_children[m_childIndex].reset();
parent->m_childOffsetsDirty = true;
parent.reset();
}
else
{
parent.reset();
for (AZ::u32 childIndex = 0; childIndex < s_numVoxelChildren; childIndex++)
{
if (m_children[childIndex])
{
m_children[childIndex]->Evict(frameDelay, forceEvict, minSize);
}
}
}
}
void Voxel::EvictGpuData(AZ::u32 frameDelay, bool forceEvict)
{
if (m_lastVisibleFrameId < (GetCurrPassMainFrameID() - frameDelay) || forceEvict)
{
if (m_block)
{
ReleaseBlock();
AZStd::shared_ptr<Voxel> parent = m_parentNode.lock();
if (parent)
{
parent->m_childOffsetsDirty = true;
}
// Force evict all children's GPU data since current voxel's was evicted
forceEvict = true;
}
}
for (AZ::u32 childIndex = 0; childIndex < s_numVoxelChildren; childIndex++)
{
if (m_children[childIndex])
{
m_children[childIndex]->EvictGpuData(frameDelay, forceEvict);
}
}
}
void Voxel::EnqueueMeshes(const EntityMeshDataMap& insertions, const EntityMeshDataMap& removals, float maxSize)
{
EntityMeshDataMap newInsertions;
EntityMeshDataMap newRemovals;
{
AZStd::lock_guard<AZStd::mutex> writeQueueLock(m_queueMutex);
AZ::u32 queueId = m_queueId;
auto& insertionQueue = m_insertions[queueId];
auto& removalQueue = m_removals[queueId];
for (auto& meshToRemove : removals)
{
//Check if mesh is in the list of inserted or pending insertion meshes
auto searchInserted = m_insertedAndPendingInsertionMeshes.find(meshToRemove.first);
if (searchInserted != m_insertedAndPendingInsertionMeshes.end())
{
newRemovals.insert(*searchInserted);
removalQueue.insert(*searchInserted);
m_insertedAndPendingInsertionMeshes.erase(searchInserted);
}
//Remove pending insertions from current queue.
auto searchInsertions = insertionQueue.find(meshToRemove.first);
if (searchInsertions != insertionQueue.end())
{
insertionQueue.erase(searchInsertions);
}
}
for (auto& meshToInsert : insertions)
{
//Check if the mesh overlaps the voxel
if (meshToInsert.second->m_worldAabb.Overlaps(m_nodeBox))
{
newInsertions.insert(meshToInsert);
bool inserted = insertionQueue.insert(meshToInsert).second;
AZ_Assert(inserted, "Tried to double insert mesh");
// Adding the mesh to the inserted + pending insertion.
inserted = m_insertedAndPendingInsertionMeshes.insert(meshToInsert).second;
AZ_Assert(inserted, "Tried to double insert mesh.");
}
}
}
// Enqueue to children new mesh insertions/removals that affected this voxel.
bool hasNewInsertionsOrRemovals = !(newInsertions.empty() && newRemovals.empty());
if (hasNewInsertionsOrRemovals)
{
for (AZ::u32 childIndex = 0; childIndex < s_numVoxelChildren; childIndex++)
{
if (m_children[childIndex])
{
m_children[childIndex]->EnqueueMeshes(newInsertions, newRemovals, maxSize);
}
}
}
}
void Voxel::UpdateBrickData(float maxSize, AZ::s32 maxLoadedNodes, DataBrick<GISubVoxels>& scratchData)
{
// If this voxel doesn't have brick data yet
// do not allocate more memory if we reached the
// maximum number of active voxels.
if (m_boxSize <= maxSize)
{
if (!m_brick || !m_brick->HasBrickData())
{
if (m_svoEnv->m_activeVoxels >= maxLoadedNodes)
{
return;
}
}
}
// Swap buffers.
AZ::u32 queueId;
{
AZStd::lock_guard<AZStd::mutex> swapQueueLock(m_queueMutex);
queueId = m_queueId;
m_queueId = (m_queueId + 1) % s_bufferCount;
}
AZ_PROFILE_FUNCTION(AZ::Debug::ProfileCategory::Renderer);
auto& insertions = m_insertions[queueId];
auto& removals = m_removals[queueId];
if (m_boxSize <= maxSize)
{
if (!m_brick)
{
m_brick = aznew Brick();
}
bool hadBrickData = m_brick->HasBrickData();
m_brick->m_brickAabb = m_nodeBox;
//Convert bounding box to local coordinates
AZ::Vector3 vCenter = m_nodeBox.GetCenter();
m_brick->m_brickAabb.Translate(-vCenter);
m_brick->m_brickOrigin = vCenter;
m_brick->ProcessMeshes(insertions, removals, scratchData);
++(m_brick->m_lastUpdated);
if (!hadBrickData && m_brick->HasBrickData())
{
++m_svoEnv->m_activeVoxels;
}
}
insertions.clear();
removals.clear();
}
AZ::Aabb Voxel::GetChildBBox(AZ::u8 childIndex)
{
AZ::u8 x = (childIndex / 4);
AZ::u8 y = (childIndex - x * 4) / 2;
AZ::u8 z = (childIndex - x * 4 - y * 2);
AZ::Vector3 vSize = m_nodeBox.GetExtents() * 0.5f;
AZ::Vector3 vOffset = vSize;
vOffset *= AZ::Vector3(x, y, z);
AZ::Aabb childBox;
childBox.SetMin(m_nodeBox.GetMin() + vOffset);
childBox.SetMax(childBox.GetMin() + vSize);
return childBox;
}
void Voxel::AllocateChildren(AZStd::shared_ptr<Voxel> self, float maxSize, float minSize)
{
//Do not allocate children if we are at minimal size.
if (m_nodeBox.GetZExtent() <= minSize)
{
return;
}
const EntityMeshDataMap noRemovals;
for (AZ::u32 childIndex = 0; childIndex < s_numVoxelChildren; ++childIndex)
{
//Check if child needs to be allocated.
if (!m_children[childIndex])
{
AZ::Aabb childBox = GetChildBBox(childIndex);
//If the child node is not visible skip its creation.
if (!m_svoEnv->m_camera.IsAABBVisible_E(AZAabbToLyAABB(childBox)))
{
continue;
}
m_children[childIndex] = AZStd::make_shared<Voxel>(childBox, self, m_svoEnv, childIndex);
//Propagate meshes to the child
if (!m_insertedAndPendingInsertionMeshes.empty())
{
m_children[childIndex]->EnqueueMeshes(m_insertedAndPendingInsertionMeshes, noRemovals, maxSize);
}
}
}
}
void Voxel::ReserveGPUMemory(AZStd::shared_ptr<Voxel> self, float maxSize)
{
if (!m_svoEnv->m_camera.IsAABBVisible_E(AZAabbToLyAABB(m_nodeBox)))
{
// If voxel is not visible, its children aren't either
return;
}
//If the voxel has data but hasn't been uploaded we will check if it needs to be uploaded.
//Note: due to how the gpu offsets are being calculated we have to allocate blocks to the larger nodes
//This will need to be fixed later.
if (!m_block && ((m_brick && m_brick->HasBrickData()) || (m_boxSize > maxSize)))
{
AZ::s32 blockWidth = 1;
AZ::s32 blockHeight = 1;
AZ::s32 blockDepth = 1;
m_blockID = m_svoEnv->m_blockPacker->AddBlock(blockWidth, blockHeight, blockDepth, m_nodeBox);
if (m_blockID != TextureBlockPacker3D::s_invalidBlockID)
{
m_block = m_svoEnv->m_blockPacker->GetBlockInfo(m_blockID);
AZ_Assert(m_block, "Invalid block ID %d", m_blockID);
}
else
{
m_svoEnv->m_evictGpu = true;
return; //Unable to reserve a block.
}
AZStd::shared_ptr<Voxel> parent = m_parentNode.lock();
if (parent)
{
parent->m_childOffsetsDirty = true;
}
}
if (m_block)
{
for (AZ::u32 childIndex = 0; childIndex < s_numVoxelChildren; childIndex++)
{
AZStd::shared_ptr<Voxel> child = m_children[childIndex];
if (child)
{
child->ReserveGPUMemory(child, maxSize);
}
}
}
}
void Voxel::UpdateGpuTree(AZStd::shared_ptr<Voxel> self)
{
// if we don't have a block on the gpu memory we should do nothing.
if (m_block)
{
// Check voxel is not already inside brick update queue
// to prevent unnecesary texture block writes.
if (!m_gpuEnqueued)
{
bool brickDataDirty = false;
if (m_brick && m_brick->m_lastUpdated != m_brick->m_lastUploaded)
{
brickDataDirty = true;
m_brick->m_lastUploaded.store(m_brick->m_lastUpdated);
}
if (m_childOffsetsDirty || brickDataDirty)
{
m_childOffsetsDirty = false;
m_svoEnv->m_brickUpdateQueue.push_back(self);
m_gpuEnqueued = true;
}
}
for (AZ::u32 childIndex = 0; childIndex < s_numVoxelChildren; childIndex++)
{
AZStd::shared_ptr<Voxel> child = m_children[childIndex];
if (child)
{
child->UpdateGpuTree(child);
}
}
}
}
AZ::s32 Voxel::GetOffset()
{
return m_block ? m_block->m_atlasOffset : -2;
}
void Voxel::UpdateTreeRenderData()
{
if (!m_block)
{
return;
}
Vec3i vOffset(m_block->m_minX, m_block->m_minY, m_block->m_minZ);
Vec4 treeData[nVoxNodMaxDim * nVoxNodMaxDim * nVoxNodMaxDim];
memset(treeData, 0x00, sizeof(treeData));
AZ::s32 childOffsets[s_numVoxelChildren] = { -2, -2, -2, -2, -2, -2, -2, -2 };
for (AZ::u32 childIndex = 0; childIndex < s_numVoxelChildren; childIndex++)
{
if (m_children[childIndex])
{
childOffsets[childIndex] = m_children[childIndex]->GetOffset();
}
}
treeData[0] = Vec4(AZVec3ToLYVec3(m_nodeBox.GetMin()), 0);
treeData[1] = treeData[0] + Vec4(Vec3(1, 1, 1) * m_nodeBox.GetZExtent(), 0);
treeData[0].w = m_nodeBox.GetZExtent();
{
AZStd::shared_ptr<Voxel> parent = m_parentNode.lock();
treeData[1].w = parent ? (0.1f + (float)parent->GetOffset()) : -2.f;
}
for (AZ::s32 c = 0; c < 4; c++)
{
if (childOffsets[c + 0] >= 0)
{
treeData[2][c] = 0.1f + (float)childOffsets[c + 0];
}
else
{
treeData[2][c] = -0.1f + (float)childOffsets[c + 0];
}
if (childOffsets[c + 4] >= 0)
{
treeData[3][c] = 0.1f + (float)childOffsets[c + 4];
}
else
{
treeData[3][c] = -0.1f + (float)childOffsets[c + 4];
}
}
treeData[4][0] = 0.1f + (float)gEnv->pRenderer->GetFrameID(false);
gEnv->pRenderer->UpdateTextureInVideoMemory(
m_svoEnv->m_nTexNodePoolId,
(AZ::u8*)&treeData[0],
vOffset.x * nVoxNodMaxDim,
vOffset.y * nVoxNodMaxDim,
nVoxNodMaxDim,
nVoxNodMaxDim,
eTF_R32G32B32A32F,
vOffset.z * nVoxNodMaxDim,
nVoxNodMaxDim);
}
void Voxel::UpdateBrickRenderData()
{
if (!m_block)
{
return;
}
if (!m_brick || !m_brick->HasBrickData())
{
return;
}
//Lock and read what data is currently there. If a job is half way through processing then it could potentially
//cause a partial upload, but the job will re-mark the data as dirty for the next possible frame.
//This is to avoid writing to the data while it is uploading.
AZStd::shared_lock<AZStd::shared_mutex> uploadLock(m_brick->m_brickDataMutex);
Vec3i vOffset(m_block->m_minX, m_block->m_minY, m_block->m_minZ);
const byte* pImgRgb = reinterpret_cast<const byte*>(m_brick->m_colors->m_data);
const byte* pImgNor = reinterpret_cast<const byte*>(m_brick->m_normals->m_data);
const byte* pImgOpa = reinterpret_cast<const byte*>(m_brick->m_opacities->m_data);
Vec3i vSizeFin;
vSizeFin.x = (brickDimension);
vSizeFin.y = (brickDimension);
vSizeFin.z = (brickDimension);
gEnv->pRenderer->UpdateTextureInVideoMemory(
m_svoEnv->m_nTexRgb0PoolId,
pImgRgb,
vOffset.x * nVoxBloMaxDim,
vOffset.y * nVoxBloMaxDim,
vSizeFin.x,
vSizeFin.y,
m_svoEnv->m_brickTextureFormat,
vOffset.z * nVoxBloMaxDim,
vSizeFin.z);
gEnv->pRenderer->UpdateTextureInVideoMemory(
m_svoEnv->m_nTexNormPoolId,
pImgNor,
vOffset.x * nVoxBloMaxDim,
vOffset.y * nVoxBloMaxDim,
vSizeFin.x,
vSizeFin.y,
m_svoEnv->m_brickTextureFormat,
vOffset.z * nVoxBloMaxDim,
vSizeFin.z);
gEnv->pRenderer->UpdateTextureInVideoMemory(
m_svoEnv->m_nTexOpasPoolId,
pImgOpa,
vOffset.x * nVoxBloMaxDim,
vOffset.y * nVoxBloMaxDim,
vSizeFin.x,
vSizeFin.y,
m_svoEnv->m_brickTextureFormat,
vOffset.z * nVoxBloMaxDim,
vSizeFin.z);
}
float Voxel::GetLodRatio()
{
const CCamera& cam = m_svoEnv->m_camera;
const float dist = m_nodeBox.GetCenter().GetDistance(LYVec3ToAZVec3(cam.GetPosition()));
return dist / m_boxSize;
}
//Debug drawing functionality.
void Voxel::DrawVoxels()
{
bool drawSelf = true;
if (GetLodRatio() > m_svoEnv->m_voxelLodCutoff)
{
drawSelf = false;
}
for (AZ::u32 childIndex = 0; childIndex < s_numVoxelChildren; childIndex++)
{
if (m_children[childIndex])
{
m_children[childIndex]->DrawVoxels();
}
}
if (drawSelf)
{
DrawBrickData();
}
}
void Voxel::DrawBrickData()
{
const CCamera& cam = m_svoEnv->m_camera;
if (cam.IsAABBVisible_F(AZAabbToLyAABB(m_nodeBox)))
{
if (m_brick && m_brick->HasBrickData())
{
AZStd::shared_lock<AZStd::shared_mutex> readLock(m_brick->m_brickDataMutex);
AZ::u8 boxLog = static_cast<AZ::u8>(log2(m_boxSize));
ColorF brickColor;
ColorF voxelColor = Col_Black;
switch ((boxLog % 3) + (m_block ? 3 : 0))
{
case 0:
brickColor = Col_Red;
break;
case 1:
brickColor = Col_Lime;
break;
case 2:
brickColor = Col_Blue;
break;
case 3:
brickColor = Col_Magenta;
break;
case 4:
brickColor = Col_Yellow;
break;
case 5:
brickColor = Col_Cyan;
break;
default:
brickColor = Col_DarkGrey;
break;
}
gEnv->pRenderer->GetIRenderAuxGeom()->SetRenderFlags(SAuxGeomRenderFlags());
gEnv->pRenderer->GetIRenderAuxGeom()->DrawAABB(AZAabbToLyAABB(m_nodeBox), false, voxelColor, eBBD_Faceted);
for (int x = 0; x < brickDimension; x++)
{
for (int y = 0; y < brickDimension; y++)
{
for (int z = 0; z < brickDimension; z++)
{
int id = z * brickDimension * brickDimension + y * brickDimension + x;
if ((*m_brick->m_counts)[id] > 0)
{
AZ::Vector3 vMin = m_nodeBox.GetMin() + (m_nodeBox.GetMax() - m_nodeBox.GetMin()) * AZ::Vector3((float)x / brickDimension, (float)y / brickDimension, (float)z / brickDimension);
AZ::Vector3 vMax = m_nodeBox.GetMin() + (m_nodeBox.GetMax() - m_nodeBox.GetMin()) * AZ::Vector3((float)(x + 1) / brickDimension, (float)(y + 1) / brickDimension, (float)(z + 1) / brickDimension);
AZ::Aabb brickBox;
brickBox.SetMin(vMin);
brickBox.SetMax(vMax);
bool doNotDraw = false;
for (AZ::u32 childIndex = 0; childIndex < s_numVoxelChildren; childIndex++)
{
Voxel* child = m_children[childIndex].get();
if (child && child->GetLodRatio() <= m_svoEnv->m_voxelLodCutoff)
{
if (Overlap::AABB_AABB(AZAabbToLyAABB(child->m_nodeBox), AZAabbToLyAABB(brickBox)))
{
doNotDraw = true;
}
}
}
if (doNotDraw)
{
continue;
}
gEnv->pRenderer->GetIRenderAuxGeom()->SetRenderFlags(SAuxGeomRenderFlags());
gEnv->pRenderer->GetIRenderAuxGeom()->DrawAABB(AZAabbToLyAABB(brickBox), false, brickColor, eBBD_Faceted);
}
}
}
}
}
}
}
} // namespace SVOGI
Reference in New Issue View Git Blame Copy Permalink