Update Spectra to v1.1.0 to fix C++20 compatibility issues on GCC/Clang

[valgur]

GTSAM currently unfortunately fails to build with C++20 or newer on GCC and Clang with the following errors in ShonanAveraging.cpp:

  [ 70%] Building CXX object gtsam/CMakeFiles/gtsam.dir/sfm/ShonanAveraging.cpp.o
  In file included from /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/SymEigsBase.h:20,
                   from /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/SymEigsSolver.h:12,
                   from /home/martin/libs/gtsam/gtsam/sfm/ShonanAveraging.cpp:19:
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/MatOp/internal/ArnoldiOp.h:108:50: error: expected ‘)’ before ‘*’ token
    108 |     ArnoldiOp<Scalar, OpType, IdentityBOp>(OpType* op, IdentityBOp* /*Bop*/) :
        |                                           ~      ^
        |                                                  )
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/SymEigsBase.h: In instantiation of ‘Spectra::SymEigsBase<Scalar, SelectionRule, OpType, BOpType>::SymEigsBase(OpType*, BOpType*, Index, Index) [with Scalar = double; int SelectionRule = 0; OpType = gtsam::MatrixProdFunctor; BOpType = Spectra::IdentityBOp; Index = long int]’:
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/SymEigsSolver.h:165:83:   required from ‘Spectra::SymEigsSolver<Scalar, SelectionRule, OpType>::SymEigsSolver(OpType*, Index, Index) [with Scalar = double; int SelectionRule = 0; OpType = gtsam::MatrixProdFunctor; Index = long int]’
  /home/martin/libs/gtsam/gtsam/sfm/ShonanAveraging.cpp:641:79:   required from here
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/SymEigsBase.h:252:15: error: no matching function for call to ‘Spectra::ArnoldiOp<double, gtsam::MatrixProdFunctor, Spectra::IdentityBOp>::ArnoldiOp(gtsam::MatrixProdFunctor*&, Spectra::IdentityBOp*&)’
    252 |         m_fac(ArnoldiOpType(op, Bop), m_ncv),
        |               ^~~~~~~~~~~~~~~~~~~~~~
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/MatOp/internal/ArnoldiOp.h:99:7: note: candidate: ‘constexpr Spectra::ArnoldiOp<double, gtsam::MatrixProdFunctor, Spectra::IdentityBOp>::ArnoldiOp(const Spectra::ArnoldiOp<double, gtsam::MatrixProdFunctor, Spectra::IdentityBOp>&)’
     99 | class ArnoldiOp<Scalar, OpType, IdentityBOp>
        |       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/MatOp/internal/ArnoldiOp.h:99:7: note:   candidate expects 1 argument, 2 provided
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/MatOp/internal/ArnoldiOp.h:99:7: note: candidate: ‘constexpr Spectra::ArnoldiOp<double, gtsam::MatrixProdFunctor, Spectra::IdentityBOp>::ArnoldiOp(Spectra::ArnoldiOp<double, gtsam::MatrixProdFunctor, Spectra::IdentityBOp>&&)’
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/MatOp/internal/ArnoldiOp.h:99:7: note:   candidate expects 1 argument, 2 provided
  gmake[2]: *** [gtsam/CMakeFiles/gtsam.dir/build.make:2050: gtsam/CMakeFiles/gtsam.dir/sfm/ShonanAveraging.cpp.o] Error 1
  gmake[1]: *** [CMakeFiles/Makefile2:1397: gtsam/CMakeFiles/gtsam.dir/all] Error 2
  gmake: *** [Makefile:166: all] Error 2

This PR updates the vendored Spectra from v0.9.0 to the latest v1.0.1 as a simple fix for these C++20 compatibility issues. I also moved the Spectra headers under a Spectra/... prefix to match the installed header paths since v0.7.0.

Change log: https://github.com/yixuan/spectra/blob/master/CHANGELOG.md

Since this is a major version bump, there are some breaking changes in the Spectra API, which are listed in the migration guide: https://spectralib.org/upgrade

Corresponding changes to ShonanAveraging.cpp:

diff --git a/gtsam/sfm/ShonanAveraging.cpp b/gtsam/sfm/ShonanAveraging.cpp
index 7c8b07f37c..151c816b21 100644
--- a/gtsam/sfm/ShonanAveraging.cpp
+++ b/gtsam/sfm/ShonanAveraging.cpp
@@ -16,7 +16,7 @@
  * @brief  Shonan Averaging algorithm
  */
 
-#include <SymEigsSolver.h>
+#include <Spectra/SymEigsSolver.h>
 #include <cmath>
 #include <gtsam/linear/PCGSolver.h>
 #include <gtsam/linear/SubgraphPreconditioner.h>
@@ -574,6 +574,8 @@ static bool PowerMinimumEigenValue(
  * nontrivial function, perform_op(x,y), that computes and returns the product
  * y = (A + sigma*I) x */
 struct MatrixProdFunctor {
+  using Scalar = double;
+
   // Const reference to an externally-held matrix whose minimum-eigenvalue we
   // want to compute
   const Sparse &A_;
@@ -634,13 +636,13 @@ static bool SparseMinimumEigenValue(
     Eigen::Index numLanczosVectors = 20) {
   // a. Estimate the largest-magnitude eigenvalue of this matrix using Lanczos
   MatrixProdFunctor lmOperator(A);
-  Spectra::SymEigsSolver<double, Spectra::SELECT_EIGENVALUE::LARGEST_MAGN,
-                         MatrixProdFunctor>
-      lmEigenValueSolver(&lmOperator, 1, std::min(numLanczosVectors, A.rows()));
+  Spectra::SymEigsSolver lmEigenValueSolver(
+      lmOperator, 1, std::min(numLanczosVectors, A.rows()));
   lmEigenValueSolver.init();
 
-  const int lmConverged = lmEigenValueSolver.compute(
-      maxIterations, 1e-4, Spectra::SELECT_EIGENVALUE::LARGEST_MAGN);
+  const int lmConverged =
+      lmEigenValueSolver.compute(Spectra::SortRule::LargestMagn, maxIterations,
+                                 1e-4, Spectra::SortRule::LargestMagn);
 
   // Check convergence and bail out if necessary
   if (lmConverged != 1) return false;
@@ -668,10 +670,8 @@ static bool SparseMinimumEigenValue(
 
   MatrixProdFunctor minShiftedOperator(A, -2 * lmEigenValue);
 
-  Spectra::SymEigsSolver<double, Spectra::SELECT_EIGENVALUE::LARGEST_MAGN,
-                         MatrixProdFunctor>
-      minEigenValueSolver(&minShiftedOperator, 1,
-                          std::min(numLanczosVectors, A.rows()));
+  Spectra::SymEigsSolver minEigenValueSolver(
+      minShiftedOperator, 1, std::min(numLanczosVectors, A.rows()));
 
   // If S is a critical point of F, then S^T is also in the null space of S -
   // Lambda(S) (cf. Lemma 6 of the tech report), and therefore its rows are
@@ -699,8 +699,9 @@ static bool SparseMinimumEigenValue(
   // order to be able to estimate the smallest eigenvalue within an *absolute*
   // tolerance of 'minEigenvalueNonnegativityTolerance'
   const int minConverged = minEigenValueSolver.compute(
-      maxIterations, minEigenvalueNonnegativityTolerance / lmEigenValue,
-      Spectra::SELECT_EIGENVALUE::LARGEST_MAGN);
+      Spectra::SortRule::LargestMagn, maxIterations,
+      minEigenvalueNonnegativityTolerance / lmEigenValue,
+      Spectra::SortRule::LargestMagn);
 
   if (minConverged != 1) return false;

To be honest, I would personally be hesitant to merge such an unwieldy PR, so you might want to commit these changes yourself or at least diff them against the official Spectra sources.

[dellaert]

@valgur sorry for the delay in reviewing. Partly the reason is it seems such a big change. Is there compelling reason to change the location of the headers? Or: can you make this PR much smaller/reviewable by using the old location? Note I have not tried to figure out the answer myself.

PS FWIW at Verdant we compiled with C++20 without any issue - so slightly confused about that - but I think upgrading Spectra is a good thing nonetheless.

[valgur]

Is there compelling reason to change the location of the headers?

Sorry, I should have indeed included my reasoning behind this. The official Spectra package uses a Spectra/ prefix for its includes since v0.7.0, e.g. #include <Spectra/SymEigsSolver.h> instead of #include <SymEigsSolver.h>, and the change in file path was to align with that. It does not really make much sense in isolation, but it's relevant for external package managers like Conan and Vcpkg, which generally try to avoid vendored libraries as much as possible and replace them with versions from the package repository. Following the conventions of the official Spectra package makes it easier to swap out the vendored version without additional patching.

That being said, it's not something that is strictly required and I can revert the move. The necessary patching is limited to just a single line anyway.

The PR will still be quite large regardless, since the update touches more than 50 files.

PS FWIW at Verdant we compiled with C++20 without any issue - so slightly confused about that - but I think upgrading Spectra is a good thing nonetheless.

The change is motivated by a GTSAM build failure on GCC 14 (if I recall correctly) with C++20 enabled, which was reported on ConanCenter Slack a couple of months ago.

I also now noticed that Spectra v1.1.0 was released just last week after a 2.5 year hiatus. 🙂 I applied the update to the PR.

[dellaert]

OK, let me take a deeper look. I think it'll be ok.

[dellaert]

Thanks !!!

I have a slight preference if the files were in the same place as before, but if this is easier for you I don't really care that much. I have some question about the mapping that I'd like to know the answer to before merging, but otherwise looks great!

[valgur]

I have a slight preference if the files were in the same place as before, but if this is easier for you I don't really care that much.

Ok, I moved the headers back to their original location.

[dellaert]

Awesome! Many thanks for contributing!
If you are using gcc14 with c++20, consider adding a special case CI flow for GTSAM. It'd be good to check for other issues as new features are developed....