QREngineResultCode GPUQREngine
(
size_t gpuMemorySize, // The total available GPU memory size in bytes
Front *userFronts, // The list of fronts to factorize
Int numFronts, // The number of fronts to factorize
QREngineStats *stats // An optional parameter. If present, statistics
// are collected and passed back to the caller
// via this struct
)
{
/* Allocate workspaces */
Front *fronts = (Front*) SuiteSparse_calloc(numFronts, sizeof(Front));
if(!fronts)
{
return QRENGINE_OUTOFMEMORY;
}
size_t FSize, RSize;
FSize = RSize = 0;
for(int f=0; f<numFronts; f++)
{
/* Configure the front */
Front *userFront = &(userFronts[f]);
Int m = userFront->fm;
Int n = userFront->fn;
Front *front = new (&fronts[f]) Front(f, EMPTY, m, n);
FSize += front->getNumFrontValues();
RSize += front->getNumRValues();
}
// We have to allocate page-locked CPU-GPU space to leverage asynchronous
// memory transfers. This has to be done in a way that the CUDA driver is
// aware of, which unfortunately means making a copy of the user input.
// calloc pagelocked space on CPU, and calloc space on the GPU
Workspace *wsMongoF = Workspace::allocate(FSize, // CPU and GPU
sizeof(double), true, true, true, true);
// calloc pagelocked space on the CPU. Nothing on the GPU
Workspace *wsMongoR = Workspace::allocate(RSize, // CPU
sizeof(double), true, true, false, true);
/* Cleanup and return if we ran out of memory. */
if(!wsMongoF || !wsMongoR)
{
return GPUQREngine_Cleanup (QRENGINE_OUTOFMEMORY,
userFronts, fronts, numFronts, wsMongoF, wsMongoR);
}
/* Prepare the fronts for GPU execution. */
size_t FOffset, ROffset;
FOffset = ROffset = 0;
for(int f=0; f<numFronts; f++)
{
// Set the front pointers; make the copy from user data into front data.
Front *front = &(fronts[f]);
front->F = CPU_REFERENCE(wsMongoF, double*) + FOffset;
front->gpuF = GPU_REFERENCE(wsMongoF, double*) + FOffset;
front->cpuR = CPU_REFERENCE(wsMongoR, double*) + ROffset;
FOffset += front->getNumFrontValues();
ROffset += front->getNumRValues();
/* COPY USER DATA (user's F to our F) */
Front *userFront = &(userFronts[f]);
double *userF = userFront->F;
double *F = front->F;
Int m = userFront->fm;
Int n = userFront->fn;
bool isColMajor = userFront->isColMajor;
Int ldn = userFront->ldn;
for(Int i=0; i<m; i++)
{
for(Int j=0; j<n; j++)
{
F[i*n+j] = (isColMajor ? userF[j*ldn+i] : userF[i*ldn+j]);
}
}
/* Attach either the user-specified Stair, or compute it. */
front->Stair = userFront->Stair;
if(!front->Stair) front->Stair = GPUQREngine_FindStaircase(front);
/* Cleanup and return if we ran out of memory building the staircase */
if(!front->Stair)
{
return GPUQREngine_Cleanup (QRENGINE_OUTOFMEMORY,
userFronts, fronts, numFronts, wsMongoF, wsMongoR);
}
}
/* Transfer the fronts to the GPU. */
if(!wsMongoF->transfer(cudaMemcpyHostToDevice))
{
return GPUQREngine_Cleanup (QRENGINE_GPUERROR,
userFronts, fronts, numFronts, wsMongoF, wsMongoR);
}
/* Do the factorization for this set of fronts. */
QREngineResultCode result = GPUQREngine_Internal(gpuMemorySize, fronts,
numFronts, NULL, NULL, NULL, stats);
if(result != QRENGINE_SUCCESS)
{
return GPUQREngine_Cleanup (result,
userFronts, fronts, numFronts, wsMongoF, wsMongoR);
}
/* COPY USER DATA (our R back to user's R) */
for(int f=0; f<numFronts; f++)
{
Front *userFront = &(userFronts[f]);
double *R = (&fronts[f])->cpuR;
double *userR = userFront->cpuR;
Int m = userFront->fm;
Int n = userFront->fn;
Int rank = userFront->rank;
bool isColMajor = userFront->isColMajor;
Int ldn = userFront->ldn;
for(Int i=0; i<rank; i++)
{
for(Int j=0; j<n; j++)
{
userR[i*ldn+j] = (isColMajor ? R[j*n+i] : R[i*n+j]);
}
}
}
/* Return that the factorization was successful. */
return GPUQREngine_Cleanup (QRENGINE_SUCCESS,
userFronts, fronts, numFronts, wsMongoF, wsMongoR);
}
EdgeCutProblem *EdgeCutProblem::create(const Int _n, const Int _nz, Int *_p,
Int *_i, double *_x, double *_w)
{
void *memoryLocation = SuiteSparse_malloc(1, sizeof(EdgeCutProblem));
if (!memoryLocation)
return NULL;
// Placement new
EdgeCutProblem *graph = new (memoryLocation) EdgeCutProblem();
graph->shallow_p = (_p != NULL);
graph->shallow_i = (_i != NULL);
graph->shallow_x = (_x != NULL);
graph->shallow_w = (_w != NULL);
size_t n = static_cast<size_t>(_n);
graph->n = _n;
size_t nz = static_cast<size_t>(_nz);
graph->nz = _nz;
graph->p = (graph->shallow_p)
? _p
: (Int *)SuiteSparse_calloc(n + 1, sizeof(Int));
graph->i
= (graph->shallow_i) ? _i : (Int *)SuiteSparse_malloc(nz, sizeof(Int));
graph->x = _x;
graph->w = _w;
graph->X = 0.0;
graph->W = 0.0;
graph->H = 0.0;
if (!graph->p || !graph->i)
{
graph->~EdgeCutProblem();
return NULL;
}
graph->partition = (bool *)SuiteSparse_malloc(n, sizeof(bool));
graph->vertexGains = (double *)SuiteSparse_malloc(n, sizeof(double));
graph->externalDegree = (Int *)SuiteSparse_calloc(n, sizeof(Int));
graph->bhIndex = (Int *)SuiteSparse_calloc(n, sizeof(Int));
graph->bhHeap[0] = (Int *)SuiteSparse_malloc(n, sizeof(Int));
graph->bhHeap[1] = (Int *)SuiteSparse_malloc(n, sizeof(Int));
graph->bhSize[0] = graph->bhSize[1] = 0;
if (!graph->partition || !graph->vertexGains || !graph->externalDegree
|| !graph->bhIndex || !graph->bhHeap[0] || !graph->bhHeap[1])
{
graph->~EdgeCutProblem();
return NULL;
}
graph->heuCost = 0.0;
graph->cutCost = 0.0;
graph->W0 = 0.0;
graph->W1 = 0.0;
graph->imbalance = 0.0;
graph->parent = NULL;
graph->clevel = 0;
graph->cn = 0;
graph->matching = (Int *)SuiteSparse_calloc(n, sizeof(Int));
graph->matchmap = (Int *)SuiteSparse_malloc(n, sizeof(Int));
graph->invmatchmap = (Int *)SuiteSparse_malloc(n, sizeof(Int));
graph->matchtype = (Int *)SuiteSparse_malloc(n, sizeof(Int));
graph->markArray = (Int *)SuiteSparse_calloc(n, sizeof(Int));
graph->markValue = 1;
graph->singleton = -1;
if (!graph->matching || !graph->matchmap || !graph->invmatchmap
|| !graph->markArray || !graph->matchtype)
{
graph->~EdgeCutProblem();
return NULL;
}
graph->initialized = false;
return graph;
}