Green's Functions #13

CarlosMejZ · 2023-06-02T15:59:11Z

These are the basic functionalities to compute one-body Green's functions for a selected CI wave function, as described in Phys. Rev. B 100, 125165 (2019) and J. Chem. Phys. 154, 121101 (2021). It implements both a continuous fraction based Lanczos and band Lanczos method. The console output needs to be adapted to match the one used in the rest of the code, and unit testing is required (so far, only a small example testing is possible within the standalone_driver.cxx test code).

…e doxygen docu.

wavefunction91

Initial round of comments - I can get things to compile locally. I'll fix CI this evening

include/macis/gf/bandlan.hpp

include/macis/gf/inn_prods.hpp

include/macis/gf/lanczos.hpp

src/macis/gf/bandlan.cxx

src/macis/gf/eigsolver.cxx

src/macis/gf/gf.cxx

wavefunction91 · 2023-06-02T19:04:15Z

@CarlosMejZ Do you have any ideas on how you might want to unit test this code? Are there exact results we can compare to / things we could reliably reproduce?

…Band Lanczos routines to use matrices which are guaranteed contiguous in memory.

src/macis/gf/bandlan.cxx

wavefunction91 · 2023-06-05T15:27:51Z

src/macis/gf/bandlan.cxx

-  delete[] WORK;
-  delete[] E;
+  D.resize(N);
+  lapack::heev_2stage(JOBZ, UPLO, N, A.data(), LDA, D.data());


Good call on the 2stage, but that' actually isn't faster for all general cases. Brings up an interesting point about how to select the algorithm though... Might be good to have a switch. I'll give it some thought.

include/macis/gf/eigsolver.hpp

include/macis/gf/gf.hpp

wavefunction91 · 2023-06-05T15:33:39Z

include/macis/gf/inn_prods.hpp

@@ -32,14 +32,7 @@ namespace macis {
 */
 inline double MyInnProd(const std::vector<double>& vecR,


At some point it might make sense to just remove these functions entirely since all overloads call the same blaspp function. Not a high-priority now, but something to keep in mind.

Would you rather have the full, lengthy blaspp call everytime, or a single wrapper to make it a little easier to read?

Globally, yes, but having local (e.g. lambda) aliases would be fine.

Ok, I can make local aliases.

wavefunction91 · 2023-06-05T15:36:09Z

src/macis/gf/bandlan.cxx

@@ -293,57 +159,42 @@ bool GetEigsysBand(std::vector<std::vector<double> > &mat, int nSupDiag,
  eigvals.clear();


This is a good piece of functionality. It might make sense to wrap it up into a nice LAPACK-style function for use elsewhere.

… all matrices to avoid using vector<vector> data types.

wavefunction91 · 2023-06-15T19:05:11Z

include/macis/gf/bandlan.hpp

-        // temp[coeff] -= bandH[it - 1][jt - 1] * qs[band_indx_j][coeff];
-        temp[coeff] -= bandH[it - 1][jt - 1] * qs[N * band_indx_j + coeff];
+        temp[coeff] -=
+            bandH[(it - 1) * nLanIts + jt - 1] * qs[N * band_indx_j + coeff];


Is there a reason you chose (what looks like to me) row-major for bandH?

Probably just the similarity with how I was linearizing qs (where each vector is stored contiguously). In the case of bandH, since it's a Hermitian matrix, it does not really matter. But I can change it if you prefer to have all matrices in column-major.

All depends on what is the most efficient storage scheme - i.e. which scheme will most likely give you the most opportunities for contiguous data access. It doesn't look like it particularly matters here.

However, it looks like this operation can be made a bit more efficient - correct me if I'm wrong, but your computing something that looks like

temp(:) -= alpha * qs(:,band_idx_j)

where alpha = bandH(it-1, jt-1). That's an axpy. Looks like it can be replaced on line 207 too

Good catch!

wavefunction91 · 2023-06-16T22:20:49Z

include/macis/gf/bandlan.hpp

-        // temp[coeff] -= bandH[it - 1][jt - 1] * qs[band_indx_j][coeff];
-        temp[coeff] -= bandH[it - 1][jt - 1] * qs[N * band_indx_j + coeff];
+        temp[coeff] -=
+            bandH[(it - 1) * nLanIts + jt - 1] * qs[N * band_indx_j + coeff];


All depends on what is the most efficient storage scheme - i.e. which scheme will most likely give you the most opportunities for contiguous data access. It doesn't look like it particularly matters here.

However, it looks like this operation can be made a bit more efficient - correct me if I'm wrong, but your computing something that looks like

temp(:) -= alpha * qs(:,band_idx_j)

where alpha = bandH(it-1, jt-1). That's an axpy. Looks like it can be replaced on line 207 too

wavefunction91 · 2023-06-16T22:21:54Z

include/macis/gf/bandlan.hpp

+ * @date 25/04/2022
+ */
+template <class Cont, class Functor>
+void MyBandLan(const Functor &H, std::vector<Cont> &qs,


Naming conventions of My<Func> should get swapped out for something more meaningful. I think BandLan works fine here.

wavefunction91 · 2023-06-16T22:23:38Z

include/macis/gf/bandlan.hpp

+                  << std::endl;
+      } else {
+#pragma omp parallel for
+        for(size_t coeff = 0; coeff < temp.size(); coeff++)


This is an axpy with alpha = 1.0/bandH(it-1,jt-1) and beta = 0

wavefunction91 · 2023-06-16T22:24:18Z

include/macis/gf/bandlan.hpp

+        bandH.resize(nLanIts * nLanIts);
+        break;
+#pragma omp parallel for
+        for(size_t coeff = 0; coeff < temp.size(); coeff++)


I'd either use scal with alpha=0 or a memset here since it's contiguous

wavefunction91 · 2023-06-16T22:26:22Z

include/macis/gf/bandlan.hpp

+
+  for(int it = 1; it <= nLanIts; it++) {
+    int band_indx_i =
+        true_indx[it];  // TO WHAT ELEMENT OF THE VECTOR SET DO WE APPLY THIS


Can you explain this indirection here? Generally we want to avoid things like this (to e.g. allow for better usage of Level-3 BLAS when possible), but if its completely necessary, it's fine.

You mean the use of the effective index true_indx? (I'm afraid I don't know what you mean by indirection)

This is what I came up with to apply the matvecs in the right order to the right vector, since in band Lanczos, unlike in the regular one, at every iteration you act on a different vector in your set of nbands. Why is this a problem for BLAS? Shouldn't it be fine as long as you pass the right memory sector to act on?

I guess that a way to avoid this true_indx variable could be to turn the for loop over it into two, one external loop going over nLanIts // nbands and one internal loop going over nbands. I think that should be completely equivalent, and then the index in the internal loop would be essentially iterating over which vector in the basis to act on with the Hamiltonian. Do you think this would make the code better?

…ze the evaluation of G[w] for each frequency in the single band case. Removed MyInnProd header, substituted by lambdas for blas::dot.

…d core space determinants into the ASCI ranking (to allow for them to be substituted).

…y models. Test case added.

…ator

… be used in GF calculation.

…ackpp was not returning the right eigenvectors, but the eigenvectors to an intermediat tridiagonal matrix.

…mespace this is.

…ls LOBPCG with random guess vectors

…ompatible with comlpex wave functions. The normal Lanczos calculation of off-diagonal elements is not appropriate for complex wave functions, though.

CarlosMejZ added 2 commits June 2, 2023 16:01

Added GF functionalities. TODO: Fix the outputs to console, polish th…

e95c43f

…e doxygen docu.

Filling in the doxygen docu for the GF-relevant routines.

dbea9ae

CarlosMejZ added the enhancement New feature or request label Jun 2, 2023

CarlosMejZ requested a review from wavefunction91 June 2, 2023 15:59

license[bot] and others added 3 commits June 2, 2023 16:00

Committing license headers

88353d3

Merge branch 'master' into feature/gf

c4704b2

Committing clang-format changes

90c234a

wavefunction91 requested changes Jun 2, 2023

View reviewed changes

David Williams-Young and others added 3 commits June 2, 2023 12:04

Remove extra omp.h inclusion

2f9d5ee

Committing clang-format changes

e23c537

Merge branch 'master' into feature/gf

a5946ab

wavefunction91 changed the title ~~Feature/gf~~ Green's Functions Jun 4, 2023

CarlosMejZ and others added 2 commits June 5, 2023 17:12

Applied review comments from pull request, started reformulating the …

2ffd1e2

…Band Lanczos routines to use matrices which are guaranteed contiguous in memory.

Committing clang-format changes

91e506b

wavefunction91 requested changes Jun 5, 2023

View reviewed changes

David Williams-Young and others added 9 commits June 7, 2023 08:56

Added Hubbard Hamiltonian generator + UT

5175a96

Committing clang-format changes

ebb7ef4

Committing license headers

84b219d

Add PBC

65f56b3

Committing clang-format changes

e0ef771

Merge branch 'master' into feature/model

9ab4627

Merge branch 'master' into feature/gf

414fdd1

Implemented suggestions for GF routines. Most importantly, vectorized…

302feef

… all matrices to avoid using vector<vector> data types.

Committing clang-format changes

a806f13

wavefunction91 reviewed Jun 15, 2023

View reviewed changes

wavefunction91 requested changes Jun 16, 2023

View reviewed changes

CarlosMejZ and others added 3 commits June 21, 2023 10:08

Cleaned GF routines from omp pragmas, only one remaining to paralleli…

f02e8b0

…ze the evaluation of G[w] for each frequency in the single band case. Removed MyInnProd header, substituted by lambdas for blas::dot.

Committing clang-format changes

a08ebf0

BUG FIXED: Passing wrong vector length to Band Lanczos. Also, include…

5baea2e

…d core space determinants into the ASCI ranking (to allow for them to be substituted).

Clang Robot and others added 17 commits September 11, 2023 13:48

Committing clang-format changes

74702c2

Included new HamiltonianGenerator, particularly efficient for impurit…

d56989c

…y models. Test case added.

Merged with feature/model, to get the single-double Hamiltonian Gener…

3241bc4

…ator

Adding clang format changes from bot.

d4a7e4a

Committing clang-format changes

3d3277d

Added option to read in vectors of ints and bools from input file, to…

41cc116

… be used in GF calculation.

Committing clang-format changes

a51b754

Fixed Bug in GF calculation: Call to band diagonalization through Lap…

ff65ee7

…ackpp was not returning the right eigenvectors, but the eigenvectors to an intermediat tridiagonal matrix.

Committing clang-format changes

b797e30

Added functionalities to specify GF frequency grid.

f838450

Committing clang-format changes

65db897

Updating calls to submdspan, since mdspan library changed in which na…

93e48e8

…mespace this is.

Committing clang-format changes

18415d5

Added option to ask for multiple eigenstates in selected_ci_diag. Cal…

e38bdbf

…ls LOBPCG with random guess vectors

Committing clang-format changes

fa2442a

Added minor modification to make Band Lanczos computation of the GF c…

b2d334e

…ompatible with comlpex wave functions. The normal Lanczos calculation of off-diagonal elements is not appropriate for complex wave functions, though.

Committing clang-format changes

365a871

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Green's Functions #13

Green's Functions #13

CarlosMejZ commented Jun 2, 2023

wavefunction91 left a comment

wavefunction91 commented Jun 2, 2023

wavefunction91 Jun 5, 2023

wavefunction91 Jun 5, 2023

CarlosMejZ Jun 13, 2023

wavefunction91 Jun 16, 2023

CarlosMejZ Jun 19, 2023

wavefunction91 Jun 5, 2023

wavefunction91 Jun 15, 2023

CarlosMejZ Jun 16, 2023

wavefunction91 Jun 16, 2023

CarlosMejZ Jun 21, 2023

wavefunction91 Jun 16, 2023

wavefunction91 Jun 16, 2023

wavefunction91 Jun 16, 2023

wavefunction91 Jun 16, 2023

wavefunction91 Jun 16, 2023

CarlosMejZ Jun 19, 2023 •

edited

Loading

		@@ -32,14 +32,7 @@ namespace macis {
		*/
		inline double MyInnProd(const std::vector<double>& vecR,

		@@ -293,57 +159,42 @@ bool GetEigsysBand(std::vector<std::vector<double> > &mat, int nSupDiag,
		eigvals.clear();

Green's Functions #13

Are you sure you want to change the base?

Green's Functions #13

Conversation

CarlosMejZ commented Jun 2, 2023

wavefunction91 left a comment

Choose a reason for hiding this comment

wavefunction91 commented Jun 2, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

CarlosMejZ Jun 19, 2023 • edited Loading

Choose a reason for hiding this comment

CarlosMejZ Jun 19, 2023 •

edited

Loading