void reduce(const Vertex rootVertex, const Vertex srcVertex, Op op, const std::vector<T_Data> sendData, std::vector<T_Data>& recvData){
static std::vector<T_Data> reduce;
static std::vector<T_Data>* rootRecvData;
static unsigned vertexCount = 0;
static bool hasRootVertex = false;
VAddr rootVAddr = locateVertex(graph, rootVertex);
VAddr srcVAddr = locateVertex(graph, srcVertex);
Context context = getGraphContext(graph);
std::vector<Vertex> vertices = getHostedVertices(graph, srcVAddr);
vertexCount++;
if(reduce.empty()){
reduce = std::vector<T_Data>(sendData.size(), 0);
}
// Reduce locally
std::transform(reduce.begin(), reduce.end(), sendData.begin(), reduce.begin(), op);
// Remember pointer of recvData from rootVertex
if(rootVertex.id == srcVertex.id){
hasRootVertex = true;
rootRecvData = &recvData;
}
// Finally start reduction
if(vertexCount == vertices.size()){
if(hasRootVertex){
cal.reduce(rootVAddr, context, op, reduce, *rootRecvData);
}
else{
cal.reduce(rootVAddr, context, op, reduce, recvData);
}
reduce.clear();
vertexCount = 0;
}
assert(vertexCount <= vertices.size());
}
/**
* @brief Returns a set of all host peer VAddrs of the *graph*
*
*/
std::vector<VAddr> getGraphHostVAddrs(Graph& graph){
std::vector<Vertex> vertices = graph.getVertices();
std::set<VAddr> vAddrs;
for(Vertex vertex : vertices){
vAddrs.insert(locateVertex(graph, vertex));
}
return std::vector<VAddr>(vAddrs.begin(), vAddrs.end());
}
void allGather(const Vertex srcVertex, T_Send sendData, T_Recv& recvData, const bool reorder){
typedef typename T_Send::value_type T_Send_Container;
static std::vector<T_Send_Container> gather;
static std::vector<T_Recv*> recvDatas;
VAddr srcVAddr = locateVertex(graph, srcVertex);
Context context = getGraphContext(graph);
std::vector<Vertex> vertices = getHostedVertices(graph, srcVAddr);
gather.insert(gather.end(), sendData.begin(), sendData.end());
recvDatas.push_back(&recvData);
if(gather.size() == vertices.size()){
std::vector<unsigned> recvCount;
cal.allGatherVar(context, gather, *(recvDatas[0]), recvCount);
if(reorder){
std::vector<typename T_Recv::value_type> recvDataReordered(recvData.size());
unsigned vAddr = 0;
for(unsigned recv_i = 0; recv_i < recvData.size(); ){
std::vector<Vertex> hostedVertices = getHostedVertices(graph, vAddr);
for(Vertex v: hostedVertices){
recvDataReordered.at(v.id) = recvDatas[0]->data()[recv_i];
recv_i++;
}
vAddr++;
}
for(unsigned i = 0; i < recvDataReordered.size(); ++i){
recvDatas[0]->data()[i] = recvDataReordered[i];
}
}
// Distribute Received Data to Hosted Vertices
//unsigned nElements = std::accumulate(recvCount.begin(), recvCount.end(), 0);
for(unsigned i = 1; i < recvDatas.size(); ++i){
std::copy(recvDatas[0]->begin(), recvDatas[0]->end(), recvDatas[i]->begin());
}
gather.clear();
}
}
void VertexCluster::calculateWeights()
{
//为了速度，尽可能不调用外部函数,少用循环
//Common::timeBegin();
int i,j,k;
int a,b,c;
D3DXVECTOR3 vab,vbc,vca;
float lab,lbc,lca;
float abc,bca,cab;
double lsum=0.0f;
for(i=0;i<mIndicesNum;i+=3)
{
a=mIndices[i];
b=mIndices[i+1];
c=mIndices[i+2];
vab=*locateVertex(b)-*locateVertex(a);
vbc=*locateVertex(c)-*locateVertex(b);
vca=*locateVertex(c)-*locateVertex(a);
//为了速度，把sqrt省了
lab=sqrt(vab.x*vab.x+vab.y*vab.y+vab.z*vab.z); lsum+=lab;
lbc=sqrt(vbc.x*vbc.x+vbc.y*vbc.y+vbc.z*vbc.z); lsum+=lbc;
lca=sqrt(vca.x*vca.x+vca.y*vca.y+vca.z*vca.z); lsum+=lca;
mEdgeLengths[a]=max(mEdgeLengths[a],lab); mEdgeLengths[b]=max(mEdgeLengths[b],lab);
mEdgeLengths[b]=max(mEdgeLengths[b],lbc); mEdgeLengths[c]=max(mEdgeLengths[c],lbc);
mEdgeLengths[c]=max(mEdgeLengths[c],lca); mEdgeLengths[a]=max(mEdgeLengths[a],lca);
vab/=lab; vbc/=lbc; vca/=lca;
abc=-vab.x*vbc.x-vab.y*vbc.y-vab.z*vbc.z;
bca=-vbc.x*vca.x-vbc.y*vca.y-vbc.z*vca.z;
cab=-vca.x*vab.x-vca.y*vab.y-vca.z*vab.z;
mAngles[b]=max(mAngles[b],abc);
mAngles[c]=max(mAngles[c],bca);
mAngles[a]=max(mAngles[a],cab);
}
mAvgEdgeLength=lsum/double(mIndicesNum);
for(i=0;i<mVerticesNum;i++)
{
mWeights[i].mID=i;
mWeights[i].mWeight=mAngles[i]*MESH_LOD_ANGLE_FACTOR+mEdgeLengths[i]*MESH_LOD_EDGE_LENGTH_FACTOR;
// Common::log("*****id ",i);
// Common::log("mAngles ",mAngles[i]);
// Common::log("mEdgeLengths ",mEdgeLengths[i]);
// Common::log("mWeight ",mWeights[i].mWeight);
}
//Common::log("mAvgEdgeLength",mAvgEdgeLength);
//Common::log("dTime", Common::timeEnd());
}
void allReduce(const Vertex srcVertex, Op op, const std::vector<T_Data> sendData, T_Recv& recvData){
static std::vector<T_Data> reduce;
static unsigned vertexCount = 0;
static std::vector<T_Recv*> recvDatas;
VAddr srcVAddr = locateVertex(graph, srcVertex);
Context context = getGraphContext(graph);
std::vector<Vertex> vertices = getHostedVertices(graph, srcVAddr);
recvDatas.push_back(&recvData);
vertexCount++;
if(reduce.empty()){
reduce = std::vector<T_Data>(sendData.size(), 0);
}
// Reduce locally
std::transform(reduce.begin(), reduce.end(), sendData.begin(), reduce.begin(), op);
// Finally start reduction
if(vertexCount == vertices.size()){
cal.allReduce(context, op, reduce, *(recvDatas[0]));
// Distribute Received Data to Hosted Vertices
for(unsigned i = 1; i < recvDatas.size(); ++i){
std::copy(recvDatas[0]->begin(), recvDatas[0]->end(), recvDatas[i]->begin());
}
reduce.clear();
vertexCount = 0;
}
assert(vertexCount <= vertices.size());
}
void gather(const Vertex rootVertex, const Vertex srcVertex, const T_Send sendData, T_Recv& recvData, const bool reorder){
typedef typename T_Send::value_type SendValueType;
typedef typename T_Recv::value_type RecvValueType;
static std::vector<SendValueType> gather;
static T_Recv* rootRecvData = NULL;
static bool peerHostsRootVertex = false;
static unsigned nGatherCalls = 0;
nGatherCalls++;
VAddr rootVAddr = locateVertex(graph, rootVertex);
Context context = getGraphContext(graph);
// Insert data of srcVertex to the end of the gather vector
gather.insert(gather.end(), sendData.begin(), sendData.end());
// Store recv pointer of rootVertex
if(srcVertex.id == rootVertex.id){
rootRecvData = &recvData;
peerHostsRootVertex = true;
}
if(nGatherCalls == hostedVertices.size()){
std::vector<unsigned> recvCount;
std::vector<unsigned> prefixsum(context.size(),0);
if(peerHostsRootVertex){
cal.gatherVar(rootVAddr, context, gather, *rootRecvData, recvCount);
// TODO
// std::partial_sum might do the job
unsigned sum = 0;
for(unsigned count_i = 0; count_i < recvCount.size(); ++count_i){
prefixsum[count_i] = sum;
sum += recvCount[count_i];
}
// Reordering code
if(reorder){
std::vector<RecvValueType> recvDataReordered(recvData.size());
for(unsigned vAddr = 0; vAddr < context.size(); vAddr++){
std::vector<Vertex> hostedVertices = getHostedVertices(graph, vAddr);
unsigned nElementsPerVertex = recvCount.at(vAddr) / hostedVertices.size();
unsigned hVertex_i=0;
for(Vertex v: hostedVertices){
std::copy(rootRecvData->begin()+(prefixsum[vAddr] + (hVertex_i * nElementsPerVertex)),
rootRecvData->begin()+(prefixsum[vAddr] + (hVertex_i * nElementsPerVertex)) + (nElementsPerVertex),
recvDataReordered.begin()+(v.id*nElementsPerVertex));
hVertex_i++;
}
}
std::copy(recvDataReordered.begin(), recvDataReordered.end(), rootRecvData->begin());
}
}
else {
cal.gatherVar(rootVAddr, context, gather, recvData, recvCount);
}
gather.clear();
nGatherCalls = 0;
}
}
Event asyncRecv(const Vertex srcVertex, const Edge edge, T& data){
VAddr srcVAddr = locateVertex(graph, srcVertex);
Context context = getGraphContext(graph);
return cal.asyncRecv(srcVAddr, edge.id, context, data);
}
inline void recv(const Vertex srcVertex, const Edge edge, T& data){
VAddr srcVAddr = locateVertex(graph, srcVertex);
Context context = getGraphContext(graph);
cal.recv(srcVAddr, edge.id, context, data);
}
Event asyncSend(const Vertex destVertex, const Edge edge, const T& data){
VAddr destVAddr = locateVertex(graph, destVertex);
Context context = getGraphContext(graph);
return cal.asyncSend(destVAddr, edge.id, context, data);
}
inline void send(const Vertex destVertex, const Edge edge, const T& data){
VAddr destVAddr = locateVertex(graph, destVertex);
Context context = getGraphContext(graph);
cal.send(destVAddr, edge.id, context, data);
}
//debug!vb也应该生成一个新的！
void VertexCluster::generateLOD(float level,ID3D10BufferPtr &vb,ID3D10BufferPtr &ib,UINT &ibLength) //mVertexValueCreated=true
{
//Common::timeBegin();
D3D10_BUFFER_DESC desc;
desc.BindFlags= D3D10_BIND_INDEX_BUFFER;
desc.CPUAccessFlags=0;
desc.MiscFlags=0;
desc.Usage=D3D10_USAGE_DEFAULT;
D3D10_SUBRESOURCE_DATA initData;
if(level<=0.0f)
{
vb=mOriginalVB;
ib=mOriginalIB;
ibLength=mIndicesNum;
return ;
}
mTolerance=getGridSize(level);
mHashTable->init(mTolerance);
//Common::log("mTolerance",mTolerance);
int i,ri;
for(i=0;i<mVerticesNum;i++)
{
ri=mWeights[i].mID;
mHashTable->addVertex(ri,locateVertex(ri),mWeights[i].mWeight);
}
ibLength=0;
int a,b,c;
int ra,rb,rc;
for(i=0;i<mIndicesNum;i+=3)
{
a=mIndices[i];
b=mIndices[i+1];
c=mIndices[i+2];
ra=mHashTable->getReplacedID(a);
rb=mHashTable->getReplacedID(b);
rc=mHashTable->getReplacedID(c);
if(ra!=rb && ra!=rc && rb!=rc)
{
mLODIB[ibLength++]=ra;
mLODIB[ibLength++]=rb;
mLODIB[ibLength++]=rc;
//Common::log("lod face index a",ra);
//Common::log("lod face index b",rb);
//Common::log("lod face index c",rc);
}
}
//for(i=0;i<mVerticesNum;i++)
// Common::log("ReplacedID",mHashTable->getReplacedID(i));
//Common::log("LOD generated! with Indices num",ibLength);
if(ibLength==0)
{
vb=NULL;
ib=NULL;
return;
}
desc.ByteWidth=sizeof(DWORD)*ibLength;
initData.pSysMem=mLODIB;
mDevice->CreateBuffer(&desc,&initData,&ib);
// for(i=0;i<ibLength;i++)
// Common::log("IB",mLODIB[i]);
desc.BindFlags=D3D10_BIND_VERTEX_BUFFER;
desc.ByteWidth=mStride*mHashTable->getCellsNum();
initData.pSysMem=mHashTable->getCellsVB();
mDevice->CreateBuffer(&desc,&initData,&vb);
// for(i=0;i<mHashTable->getCellsNum();i++)
// Common::log("LOD VB",*mHashTable->locateVertex(i));
//Common::log("LOD generating time",Common::timeEnd());
}