portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-11 03:46:55 +01:00
Code Releases Activity

Staggered multinode block cg debugged. Missing global sum.

Browse Source 
Code stalls and resumes on KNL at cambridge. Curious.

CG iterations 23ms each, then 3200 ms pauses. Mean bandwidth reports
as 200MB/s. Comms dominant in the report. However, the time behaviour suggests it
is *bursty*.... Could be swap to disk?
This commit is contained in:
Azusa Yamaguchi
2017-08-23 15:07:18 +01:00
parent 383ca7d392
commit d9cd4f0273
5 changed files with 110 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
lib/algorithms/iterative/BlockConjugateGradient.h
View File

@ -199,7 +199,12 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
Linop.HermOp(X, AD);
tmp = B - AD;
//std::cout << GridLogMessage << " initial tmp " << norm2(tmp)<< std::endl;
ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
//std::cout << GridLogMessage << " initial Q " << norm2(Q)<< std::endl;
//std::cout << GridLogMessage << " m_rr " << m_rr<<std::endl;
//std::cout << GridLogMessage << " m_C " << m_C<<std::endl;
//std::cout << GridLogMessage << " m_Cinv " << m_Cinv<<std::endl;
D=Q;
std::cout << GridLogMessage<<"BlockCGrQ computed initial residual and QR fact " <<std::endl;
@ -221,13 +226,15 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
MatrixTimer.Start();
Linop.HermOp(D, Z);
MatrixTimer.Stop();
//std::cout << GridLogMessage << " norm2 Z " <<norm2(Z)<<std::endl;
//4. M = [D^dag Z]^{-1}
sliceInnerTimer.Start();
sliceInnerProductMatrix(m_DZ,D,Z,Orthog);
sliceInnerTimer.Stop();
m_M = m_DZ.inverse();
//std::cout << GridLogMessage << " m_DZ " <<m_DZ<<std::endl;
//5. X = X + D MC
m_tmp = m_M * m_C;
sliceMaddTimer.Start();