Fix HyKKT solver reuse and add export support by tamar-dewilde · Pull Request #464 · ORNL/ReSolve

tamar-dewilde · 2026-08-01T00:36:08Z

Description

This PR exports the HyKKT libraries and fixes issues found while reusing HyKKTSolver through the HiOp integration.

The existing HyKKT tests primarily covered a single solve. Repeated solves with the same sparsity pattern and updated numerical values exposed stale matrix data, cached solver state, repeated allocations, undersized GPU transpose workspaces, and zero-residual handling issues.

@shakedregev

Proposed changes

Export the HyKKT libraries through the installed ReSolve package so downstream packages can link them.
Refresh numerical data when HyKKTSolver is reused:
- Copy updated matrix values into the existing CHOLMOD matrix before numerical factorization.
- Reload the SpGEMM matrix inputs before each computation.
- Refresh the external D_s values pointer before each solve.
- Update the SpGEMM coefficient when gamma changes.
Fix CUDA SpGEMM output state. A local C_nnz_ variable shadowed the class member, so the stored output nonzero count was not updated.
Resize CUDA and HIP transpose workspaces when a later transpose requires a larger buffer. Reuse the existing allocation when it is large enough and grow it when necessary.
Reuse HyKKT allocations:
- Allocate the SCCG solver and its work vectors once instead of replacing them on every solve.
- Reuse the existing H_tilde_ allocation when solving without J_d.
- Reset the SCCG work vectors and scalar state before each solve while reusing their allocations.
Remove redundant device synchronizations from the repeated solve path.
Preserve the initial J_d state and reject reuse when J_d changes between empty and nonempty. Reuse still requires unchanged sparsity patterns.
Handle zero residuals:
- Return from SCCG when the initial residual is zero to avoid division by zero.
- Only normalize the final residual when the RHS norm is nonzero.
- Return an absolute residual when the RHS norm is zero.
Add regression coverage for:
- Solver reuse after refreshing J and RHS data, replacing and updating D_s, and changing gamma.
- Repeated solves that start without J_d.
- Rejected reuse when J_d changes from nonempty to empty.
- Exact zero SCCG residuals and zero RHS values.
- Reusing one transpose workspace for a small transpose followed by a much larger transpose.

Checklist

Further comments

The transpose workspace change keeps the current workspace design and grows the CUDA and HIP transpose buffers when needed rather than implementing the broader redesign proposed in #345.

andrewxu319

I can't run it right now, but I left comments.

andrewxu319 · 2026-08-01T16:17:59Z

-                << ", allocated = " << allocated_
-                << "\n";
+      std::cout << "ERROR: Changing J_d between empty and nonempty is not "
+                   "supported when reusing HyKKT.\n";


Why not put this check directly inside setMatrixBlocks()?

Also, I'm pretty sure that if J_d is empty but J_d_flag_ is true, everything would still work. So if on the first solve J_d is nonempty, then on the second solve J_d is empty, you can keep J_d_flag_ = true and maybe throw a warning about algorithm inefficiency.

setMatrixBlocks() already performs the compatibility check and records the result in status_. Since it returns void, solve() is where the error is returned.

I tried keeping J_d_flag_ = true after replacing a nonempty J_d with an empty one. The next solve segfaults in Sparse::copyValues(). Reading that path, J_d_scaled_ still has the original nonempty structure while the source is now empty.

I've left the empty/nonempty J_d transition unsupported here. The proposed setup() refactor would be the right place to rebuild the cached state that would make it work.

solve() returns real_type, so it seems strange to return 1 as an error code. I was suggesting putting the error message directly inside setMatrixBlocks(), and if you want to return something, it makes more sense to return it as an int in setMatrixBlocks().

I had kept the existing status_ flow because setMatrixBlocks() returned void, so the compatibility result was carried into solve(). I understand now that you were suggesting changing the setter API itself.

setMatrixBlocks() now returns an int and performs the check before updating the stored pointers, so a rejected call leaves the solver in its previous valid state. The error is reported through out::error() rather than std::cout. I removed status_ and the guard from solve(), and the test now checks the setter result directly.

andrewxu319 · 2026-08-01T16:22:20Z

Suggested change

Remove these two lines because they're redundant with 413-414

Removed them. I also moved loadResultMatrix() into computeSpGEMMHgamma(), since the HIP implementation needs the product and sum dimensions before it initializes the result descriptor.

andrewxu319 · 2026-08-01T16:25:31Z

    matrixHandler_->matvec(J_perm_, omega_perm_, schur_, &ONE, &MINUS_ONE, memspace_);

-    sccg_ = new SchurComplementConjugateGradient(J_->getNumRows(), J_->getNumColumns(), cholesky_, matrixHandler_, vectorHandler_, memspace_);
+    if (!allocated_)


Move if (!allocated_) to inside hykkt::HyKKTSolver::solve() (line 207) for consistency.

andrewxu319 · 2026-08-01T16:32:23Z

+      }

      y_->setToZero(memspace_);
      z_->setToZero(memspace_);


I think you can move lines 106-115 to SchurComplementConjugateGradient::solve() and remove if (!y_). Nobody is going to call this multiple times on the same RHS. Probably doesn't matter that much, but I feel like that's more logical because otherwise the user needs to call setup() before every solve().

andrewxu319 · 2026-08-01T16:39:57Z


+      void setAlpha(real_type alpha)
+      {
+        impl_->setAlpha(alpha);


Move this to SpGEMM.cpp

Same for SpGEMMCpu.hpp, etc

Also, might as well rename it setCoefficients or setConstants and let it set both alpha and beta

Renamed it to setCoefficients(alpha, beta) and updated all three backends to set both values. I added a test that changes beta and checks that the result changes on CPU, CUDA, and HIP. I also removed beta_ = beta from the CPU backend to make sure the test catches it, and it failed as expected.

andrewxu319 · 2026-08-01T16:59:13Z

        status *= validateResult(error, tol);

+        // Replace D_s and restore data modified by the first solve.
+        std::ifstream D_s_reuse_file(D_s_file_name);


Why make these reuse matrices/vectors instead of just modifying the data in J, r_x, r_s, etc?

J, r_x, r_s, r_y, and r_yd now update in place instead of creating separate reuse copies. I kept D_s as a replacement because the fix refreshes its cached values pointer. Updating the same object would not test that path.

andrewxu319 · 2026-08-01T17:01:47Z

+        real_type second_error = hykktSolver.solve();
+        status *= validateResult(second_error, tol);
+
+        // Verify an exact zero RHS is handled without producing NaNs.


Suggested change

// Verify an exact zero RHS is handled without producing NaNs.

// Check that a zero RHS doesn't result in NaNs.

andrewxu319 · 2026-08-01T17:03:27Z

+        real_type zero_rhs_error = hykktSolver.solve();
+        status *= validateResult(zero_rhs_error, tol);
+
+        // Changing J_d between nonempty and empty invalidates cached solver data.


Suggested change

// Changing J_d between nonempty and empty invalidates cached solver data.

// Check that the solver raises an error when trying to change J_d from nonempty to empty.

andrewxu319 · 2026-08-01T17:07:02Z

Suggested change

* @pre J_d_flag_ determines if variables used for Spgemm H_tilde

Applied. I reworded it slightly for clarity.

andrewxu319 · 2026-08-01T17:11:24Z

-    status_   = true; // when using API, we can't check if sparsity pattern changed
-    J_d_flag_ = J_d_flag;
+    // Arbitrary sparsity changes remain the caller's responsibility, but
+    // switching between empty and nonempty J_d invalidates cached HyKKT data.


Probably for another PR, but I think we should make a separate HyKKTSolver::setup() function and remove the allocated_ flag. This matches other parts of the codebase better. This also allows switching J_d_ from nonempty to empty as long as the user calls setup() first.

This also means you don't need to make a new solver at line 224 of HyKKTSolverTests.hpp, for example.

Agreed. I think that makes sense as a follow-up. For now, I kept the separate solver in the test so the empty J_d case still covers initialization and reuse.

Never mind, not possible. Some of the "setup" steps depend on the results of earlier "compute" steps.

tamar-dewilde · 2026-08-03T03:03:35Z

Thanks for the review @andrewxu319 I appreciate it! I made the SCCG lifecycle, SpGEMM, and reuse-test changes. All tests are passing on my end.

shakedregev

Looks good, testing.

tamar-dewilde · 2026-08-03T16:28:44Z

Looks good, testing.

The branch has been updated since this comment to move the matrix compatibility error handling into setMatrixBlocks(). Please use the latest version for testing.

shakedregev

Looks good and works, but fix these warnings.

/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.hpp:21:12: warning: 'loadProductMatrices' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   21 |       void loadProductMatrices(matrix::Csr* A, matrix::Csr* B);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:28:20: note: overridden virtual function is here
   28 |       virtual void loadProductMatrices(matrix::Csr* A, matrix::Csr* B) = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.cpp:7:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.hpp:22:12: warning: 'loadSumMatrix' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   22 |       void loadSumMatrix(matrix::Csr* D);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:29:20: note: overridden virtual function is here
   29 |       virtual void loadSumMatrix(matrix::Csr* D)                       = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.cpp:7:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.hpp:23:12: warning: 'loadResultMatrix' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   23 |       void loadResultMatrix(matrix::Csr** E_ptr);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:30:20: note: overridden virtual function is here
   30 |       virtual void loadResultMatrix(matrix::Csr** E_ptr)               = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.cpp:7:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.hpp:25:12: warning: 'compute' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   25 |       void compute();
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:32:20: note: overridden virtual function is here
   32 |       virtual void compute() = 0;
      |                    ^
4 warnings generated when compiling for gfx90a.
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMM.cpp:8:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.hpp:21:12: warning: 'loadProductMatrices' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   21 |       void loadProductMatrices(matrix::Csr* A, matrix::Csr* B);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:28:20: note: overridden virtual function is here
   28 |       virtual void loadProductMatrices(matrix::Csr* A, matrix::Csr* B) = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMM.cpp:8:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.hpp:22:12: warning: 'loadSumMatrix' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   22 |       void loadSumMatrix(matrix::Csr* D);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:29:20: note: overridden virtual function is here
   29 |       virtual void loadSumMatrix(matrix::Csr* D)                       = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMM.cpp:8:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.hpp:23:12: warning: 'loadResultMatrix' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   23 |       void loadResultMatrix(matrix::Csr** E_ptr);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:30:20: note: overridden virtual function is here
   30 |       virtual void loadResultMatrix(matrix::Csr** E_ptr)               = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMM.cpp:8:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.hpp:25:12: warning: 'compute' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   25 |       void compute();
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:32:20: note: overridden virtual function is here
   32 |       virtual void compute() = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMM.cpp:12:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMHip.hpp:27:12: warning: 'loadProductMatrices' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   27 |       void loadProductMatrices(matrix::Csr* A, matrix::Csr* B);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:28:20: note: overridden virtual function is here
   28 |       virtual void loadProductMatrices(matrix::Csr* A, matrix::Csr* B) = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMM.cpp:12:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMHip.hpp:28:12: warning: 'loadSumMatrix' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   28 |       void loadSumMatrix(matrix::Csr* D);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:29:20: note: overridden virtual function is here
   29 |       virtual void loadSumMatrix(matrix::Csr* D)                       = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMM.cpp:12:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMHip.hpp:29:12: warning: 'loadResultMatrix' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   29 |       void loadResultMatrix(matrix::Csr** E_ptr);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:30:20: note: overridden virtual function is here
   30 |       virtual void loadResultMatrix(matrix::Csr** E_ptr)               = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMM.cpp:12:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMHip.hpp:31:12: warning: 'compute' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   31 |       void compute();
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:32:20: note: overridden virtual function is here
   32 |       virtual void compute() = 0;
      |                    ^
8 warnings generated when compiling for gfx90a.
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMHip.cpp:7:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMHip.hpp:27:12: warning: 'loadProductMatrices' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   27 |       void loadProductMatrices(matrix::Csr* A, matrix::Csr* B);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:28:20: note: overridden virtual function is here
   28 |       virtual void loadProductMatrices(matrix::Csr* A, matrix::Csr* B) = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMHip.cpp:7:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMHip.hpp:28:12: warning: 'loadSumMatrix' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   28 |       void loadSumMatrix(matrix::Csr* D);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:29:20: note: overridden virtual function is here
   29 |       virtual void loadSumMatrix(matrix::Csr* D)                       = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMHip.cpp:7:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMHip.hpp:29:12: warning: 'loadResultMatrix' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   29 |       void loadResultMatrix(matrix::Csr** E_ptr);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:30:20: note: overridden virtual function is here
   30 |       virtual void loadResultMatrix(matrix::Csr** E_ptr)               = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMHip.cpp:7:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMHip.hpp:31:12: warning: 'compute' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   31 |       void compute();
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:32:20: note: overridden virtual function is here
   32 |       virtual void compute() = 0;
      |                    ^
4 warnings generated when compiling for gfx90a.
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.cpp:7:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.hpp:21:12: warning: 'loadProductMatrices' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   21 |       void loadProductMatrices(matrix::Csr* A, matrix::Csr* B);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:28:20: note: overridden virtual function is here
   28 |       virtual void loadProductMatrices(matrix::Csr* A, matrix::Csr* B) = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.cpp:7:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.hpp:22:12: warning: 'loadSumMatrix' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   22 |       void loadSumMatrix(matrix::Csr* D);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:29:20: note: overridden virtual function is here
   29 |       virtual void loadSumMatrix(matrix::Csr* D)                       = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.cpp:7:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.hpp:23:12: warning: 'loadResultMatrix' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   23 |       void loadResultMatrix(matrix::Csr** E_ptr);
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:30:20: note: overridden virtual function is here
   30 |       virtual void loadResultMatrix(matrix::Csr** E_ptr)               = 0;
      |                    ^
In file included from /ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.cpp:7:
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMCpu.hpp:25:12: warning: 'compute' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
   25 |       void compute();
      |            ^
/ccs/home/regevs/ReSolve_dir/ReSolve/resolve/hykkt/spgemm/SpGEMMImpl.hpp:32:20: note: overridden virtual function is here
   32 |       virtual void compute() = 0;

tamar-dewilde · 2026-08-03T17:18:16Z

Looks good and works, but fix these warnings.

Fixed the missing override specifiers in the CPU and HIP SpGEMM declarations. I reran the CPU, CUDA, and HIP warning builds and tests, along with the HiOp HyKKT tests. Everything passed, and the warnings are gone.

shakedregev

Great work

Stale

tamar-dewilde added 14 commits July 27, 2026 22:39

Export HyKKT libraries

189e6e0

Fix HyKKT numerical updates on solver reuse

5bc050b

Fix CUDA SpGEMM nonzero count state

a4eccb2

Resize GPU transpose workspaces as needed

8bf63dc

Remove redundant HyKKT device synchronizations

bc8e955

Reuse HyKKT conjugate gradient solver

fad8ee4

Reject incompatible HyKKT solver reuse

edb239d

Reuse HyKKT matrix allocation without J_d

fe27691

Refresh HyKKT SpGEMM gamma on solver reuse

7230464

Refresh HyKKT D_s data on solver reuse

eaa2c80

Reuse HyKKT SCCG work vectors

10ea614

Handle zero residuals in HyKKT solves

fc7c3c7

Clarify HyKKT reuse checks and residual reporting

b9916c2

Update CHANGELOG.md

b5f773e

tamar-dewilde requested review from andrewxu319, pelesh and shakedregev August 1, 2026 00:36

tamar-dewilde self-assigned this Aug 1, 2026

tamar-dewilde added bug Something isn't working enhancement New feature or request hip cuda labels Aug 1, 2026

tamar-dewilde changed the title ~~Tamar/hykkt integration fixes~~ Fix HyKKT solver reuse and add export support Aug 1, 2026

tamar-dewilde mentioned this pull request Aug 1, 2026

improve error handling and propagation in resolve #344

Open

andrewxu319 previously requested changes Aug 2, 2026

View reviewed changes

tamar-dewilde added 5 commits August 2, 2026 18:32

Clean up HyKKT SpGEMM setup

7de350b

Move HyKKT SCCG allocation into solve

fe5fd93

Reset HyKKT SCCG state in solve

b5a0b04

Set HyKKT SpGEMM coefficients

b1e5f22

Update HyKKT reuse inputs in place

e280603

Apply pre-commmit fixes

1ff0ea7

shakedregev reviewed Aug 3, 2026

View reviewed changes

Report HyKKT block errors from setMatrixBlocks

b782d0b

shakedregev requested changes Aug 3, 2026

View reviewed changes

Fix SpGEMM override warnings

70234c0

shakedregev approved these changes Aug 3, 2026

View reviewed changes

shakedregev merged commit 6b14bd4 into develop Aug 3, 2026
6 checks passed

	// Verify an exact zero RHS is handled without producing NaNs.
	// Check that a zero RHS doesn't result in NaNs.

	// Changing J_d between nonempty and empty invalidates cached solver data.
	// Check that the solver raises an error when trying to change J_d from nonempty to empty.


	* @pre J_d_flag_ determines if variables used for Spgemm H_tilde

Uh oh!

Conversation

tamar-dewilde commented Aug 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Proposed changes

Checklist

Further comments

Uh oh!

andrewxu319 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tamar-dewilde commented Aug 3, 2026

Uh oh!

shakedregev left a comment

Choose a reason for hiding this comment

Uh oh!

tamar-dewilde commented Aug 3, 2026

Uh oh!

shakedregev left a comment

Choose a reason for hiding this comment

Uh oh!

tamar-dewilde commented Aug 3, 2026

Uh oh!

shakedregev left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

tamar-dewilde commented Aug 1, 2026 •

edited

Loading