slightly update Lanczos to make it identical to v1.3.1

Author: brunowu

ChASE-MPI/impl/chase_mpidla.hpp

@@ -897,6 +897,7 @@ class ChaseMpiDLA : public ChaseMpiDLAInterface<T>
         nvtxRangePop();
         nvtxRangePushA("allreduce");
 #endif
+
         AllReduce(allreduce_backend, rsd + locked, unconverged,
                   getMPI_Type<Base<T>>(), MPI_SUM, row_comm_, mpi_wrapper_);
         //  Base<T> *resid_h;
@@ -1334,6 +1335,7 @@ class ChaseMpiDLA : public ChaseMpiDLAInterface<T>
                 nvtxRangePop();
         #endif
         */
+        std::cout << C[0] << " " << ritzVc[0] << std::endl;
         dla_->LanczosDos(idx, m, ritzVc);
     }
 
@@ -1415,6 +1417,9 @@ class ChaseMpiDLA : public ChaseMpiDLAInterface<T>
             std::swap(v1_, v0_);
             std::swap(v1_, v2_);
         }
+
+        Memcpy(memcpy_mode[2], C, v1_, m_ * sizeof(T));
+
 #ifdef USE_NSIGHT
         nvtxRangePop();
 #endif

ChASE-MPI/impl/chase_mpidla_blaslapack.hpp

@@ -310,6 +310,7 @@ class ChaseMpiDLABlaslapack : public ChaseMpiDLAInterface<T>
         t_gemm(CblasColMajor, CblasNoTrans, CblasNoTrans, m_, idx, m, &alpha,
                C_, m_, ritzVc, m, &beta, C2_, m_);
         std::memcpy(C_, C2_, m * m_ * sizeof(T));
+
     }
     void Lanczos(std::size_t M, int idx, Base<T>* d, Base<T>* e,
                  Base<T>* r_beta) override