wztzjhn / quantum_basis Goto Github PK

partial and selective re-orthogonalization for home-made Lanczos with 3 vectors.

DGKS for home-made IRAM with ncv vectors.

motivated by Lauchli's chapter in Lacroix's book, similar to the divide and conquer idea.

The matrix storage can be distributed among different nodes by using PETSc, making the ~50 spin-1/2 problems feasible.

https://www.mcs.anl.gov/petsc/index.html

1. spin-1/2 xxz chain
2. spin-1 xxz chain
3. Kondo lattice chain
4. Hubbard square lattice
5. t-J triangular lattice

in each case, use either no symmetry or with translational symmetry

may be proportional to mpi threads?

need complete the implementation of mopr sum and product

possible?

the basis with bits is not a good representation for dilute problems.

Motivated by HPhi project:
http://qlms.github.io/HPhi/index_en.html

when mkl distributed csr matrices support matrix-vector product, we can use it to parallelize the code on distributed machines.

http://www.ecs.umass.edu/~polizzi/feast/index.htm

otherwise too dangerous...

1. For single thread, the speed is slower than the binary generated directly by makefile
2. For openmp, some machine scales correctly; some even slows down. At the same time, binary generated directly by makefile scales correctly.

boost version costing too much memory

when changing schedule of one loop, another loop executed much earlier changes speed, unbelievable...

For the following pattern, Lin Table failed to build. Need smarter labeling.
x x x
x o x
x x x
(horizontal axis for ia, vertical axis for ib. "o" represents no basis here)

Also add locking and purging, to enhance stability.

Currently using bubble sort to count the "-" sign generated by lattice transformation. Need replace with quick sort in the future.