Fix two 'make test' failures in a CUDA build using HYPRE+CUDA.

mfem · Mar 19, 2024 · 1314428 · 1314428
1 parent e7400c2
commit 1314428
Show file tree

Hide file tree

Showing 2 changed files with 18 additions and 8 deletions.
diff --git a/fem/fespace.cpp b/fem/fespace.cpp
@@ -606,13 +606,15 @@ void FiniteElementSpace::GetEssentialTrueDofs(const Array<int> &bdr_attr_is_ess,
 
       int counter = 0;
       std::string error_msg = "failed dof: ";
+      auto ess_tdofs_ = ess_tdofs.HostRead();
+      auto ess_tdofs2_ = ess_tdofs2.HostRead();
       for (int i = 0; i < ess_tdofs2.Size(); ++i)
       {
-         if (bool(ess_tdofs[i]) != bool(ess_tdofs2[i]))
+         if (bool(ess_tdofs_[i]) != bool(ess_tdofs2_[i]))
          {
             error_msg += std::to_string(i) += "(R ";
-            error_msg += std::to_string(bool(ess_tdofs[i])) += " P^T ";
-            error_msg += std::to_string(bool(ess_tdofs2[i])) += ") ";
+            error_msg += std::to_string(bool(ess_tdofs_[i])) += " P^T ";
+            error_msg += std::to_string(bool(ess_tdofs2_[i])) += ") ";
             counter++;
          }
       }

diff --git a/linalg/sparsemat.cpp b/linalg/sparsemat.cpp
@@ -983,14 +983,22 @@ void SparseMatrix::AddMultTranspose(const Vector &x, Vector &y,
    }
    else
    {
+      real_t *yp = y.HostReadWrite();
+      const real_t *xp = x.HostRead();
+
+      const int *Ip = HostRead(I, height+1);
+      const int nnz = Ip[height];
+      const int *Jp = HostRead(J, nnz);
+      const real_t *Ap = HostRead(A, nnz);
+
       for (int i = 0; i < height; i++)
       {
-         const real_t xi = a * x[i];
-         const int end = I[i+1];
-         for (int j = I[i]; j < end; j++)
+         const real_t xi = a * xp[i];
+         const int end = Ip[i+1];
+         for (int j = Ip[i]; j < end; j++)
          {
-            const int Jj = J[j];
-            y[Jj] += A[j] * xi;
+            const int Jj = Jp[j];
+            yp[Jj] += Ap[j] * xi;
          }
       }
    }