void ConjugateGradientSolver::solveGpu(Vector3F * X, CudaCSRMatrix * stiffnessMatrix)
{
void * dFixed = m_deviceIsFixed->bufferOnDevice();
m_deviceX->hostToDevice(X);
void * dX = m_deviceX->bufferOnDevice();
m_deviceRhs->hostToDevice(m_b);
void * dRhs = m_deviceRhs->bufferOnDevice();
cuConjugateGradient_prevresidual((float3 *)m_devicePrev->bufferOnDevice(),
(float3 *)m_deviceResidual->bufferOnDevice(),
(mat33 *)stiffnessMatrix->deviceValue(),
(uint *)stiffnessMatrix->deviceRowPtr(),
(uint *)stiffnessMatrix->deviceColInd(),
(uint *)dFixed,
(float3 *)dX,
(float3 *)dRhs,
m_numRows);
// ceglg.write("cg init");
// ceglg.writeVec3(m_deviceResidual, m_numRows, "cg residual", CudaDbgLog::FAlways);
// ceglg.writeVec3(m_devicePrev, m_numRows, "cg prev", CudaDbgLog::FAlways);
// m_devicePrev->deviceToHost(m_prev->data());
Vector3F * prev = (Vector3F *)m_prev->data();
for(unsigned k=0; k< 1; k++) {
// std::cout<<" prev["<<k<<"] h "<<prev[k];
}
for(unsigned i=0;i<i_max;i++) {
cuConjugateGradient_Ax((float3 *)m_devicePrev->bufferOnDevice(),
(float3 *)m_deviceUpdate->bufferOnDevice(),
(float3 *)m_deviceResidual->bufferOnDevice(),
(float *)m_deviceD->bufferOnDevice(),
(float *)m_deviceD2->bufferOnDevice(),
(mat33 *)stiffnessMatrix->deviceValue(),
(uint *)stiffnessMatrix->deviceRowPtr(),
(uint *)stiffnessMatrix->deviceColInd(),
(uint *)dFixed,
m_numRows);
// ceglg.write("cg step ");
// ceglg.write(i);
// ceglg.writeVec3(m_deviceUpdate, m_numRows, "cg update", CudaDbgLog::FAlways);
// ceglg.writeVec3(m_deviceResidual, m_numRows, "cg residual", CudaDbgLog::FAlways);
// ceglg.writeVec3(m_devicePrev, m_numRows, "cg prev", CudaDbgLog::FAlways);
float d =0;
float d2=0;
m_reduce->sumF(d, (float *)m_deviceD->bufferOnDevice(), m_numRows);
m_reduce->sumF(d2, (float *)m_deviceD2->bufferOnDevice(), m_numRows);
if(fabs(d2)< 1e-10f)
d2 = 1e-10f;
float d3 = d/d2;
cuConjugateGradient_addX((float3 *)dX,
(float3 *)m_deviceResidual->bufferOnDevice(),
(float *)m_deviceD->bufferOnDevice(),
(float3 *)m_devicePrev->bufferOnDevice(),
(float3 *)m_deviceUpdate->bufferOnDevice(),
d3,
(uint *)dFixed,
m_numRows);
// ceglg.write("addX");
// ceglg.writeVec3(m_deviceUpdate, m_numRows, "cg update", CudaDbgLog::FAlways);
// ceglg.writeVec3(m_deviceResidual, m_numRows, "cg residual", CudaDbgLog::FAlways);
// ceglg.writeVec3(m_devicePrev, m_numRows, "cg prev", CudaDbgLog::FAlways);
//
float d1 = 0.f;
m_reduce->sumF(d1, (float *)m_deviceD->bufferOnDevice(), m_numRows);
// if(i>29) std::cout<<" d1["<<i<<"] "<<d1<<" ";
if(i >= i_max && d1 < 0.001f)
break;
if(fabs(d)<1e-10f)
d = 1e-10f;
float d4 = d1/d;
cuConjugateGradient_addResidual((float3 *)m_devicePrev->bufferOnDevice(),
(float3 *)m_deviceResidual->bufferOnDevice(),
d4,
(uint *)dFixed,
m_numRows);
// ceglg.write("add residual");
// ceglg.writeVec3(m_deviceUpdate, m_numRows, "cg update", CudaDbgLog::FAlways);
// ceglg.writeVec3(m_deviceResidual, m_numRows, "cg residual", CudaDbgLog::FAlways);
// ceglg.writeVec3(m_devicePrev, m_numRows, "cg prev", CudaDbgLog::FAlways);
//
}
cudaThreadSynchronize();
m_deviceX->deviceToHost(X);
}
void CudaConjugateGradientSolver::solve(void * X,
CudaCSRMatrix * A,
void * fixed,
float * error)
{
// cglg.writeVec3(m_rhs, m_dimension, "cg b", CudaDbgLog::FAlways);
//cglg.writeMat33(A->valueBuf(),
// A->numNonZero(),
// " cg A ", CudaDbgLog::FAlways);
cuConjugateGradient_prevresidual((float3 *)previous(),
(float3 *)residual(),
(mat33 *)A->deviceValue(),
(uint *)A->deviceRowPtr(),
(uint *)A->deviceColInd(),
(uint *)fixed,
(float3 *)X,
(float3 *)rightHandSide(),
m_dimension);
for(int i=0; i<FemGlobal::CGSolverMaxNumIterations; i++) {
cuConjugateGradient_Ax((float3 *)previous(),
(float3 *)updated(),
(float3 *)residual(),
(float *)diff(),
(float *)diff2(),
(mat33 *)A->deviceValue(),
(uint *)A->deviceRowPtr(),
(uint *)A->deviceColInd(),
(uint *)fixed,
m_dimension);
float d =0;
float d2=0;
m_reduce->sum<float>(d, (float *)diff(), m_dimension);
m_reduce->sum<float>(d2, (float *)diff2(), m_dimension);
if(fabs(d2)< 1e-10f)
d2 = 1e-10f;
float d3 = d/d2;
cuConjugateGradient_addX((float3 *)X,
(float3 *)residual(),
(float *)diff(),
(float3 *)previous(),
(float3 *)updated(),
d3,
(uint *)fixed,
m_dimension);
float d1 = 0.f;
m_reduce->sum<float>(d1, (float *)diff(), m_dimension);
if(error) *error = d1;
if(d1 < 0.01f)
break;
if(fabs(d)<1e-10f)
d = 1e-10f;
float d4 = d1/d;
cuConjugateGradient_addResidual((float3 *)previous(),
(float3 *)residual(),
d4,
(uint *)fixed,
m_dimension);
}
}
