OpenMP fixes for NVHPC

Author: pghysels

src/BLR/BLRMatrix.GPU.cpp

@@ -216,15 +216,21 @@ namespace strumpack {
             b3.run(scalar_t(-1.), scalar_t(1.), comp_stream, handle);
           }
           if (i > 0) {
+#if defined(STRUMPACK_USE_OPENMP_TASKLOOP)
 #pragma omp taskloop
+#endif
               for (std::size_t j=0; j<rb; j++)
                 B11.tile(j, i-1).move_to_cpu
                   (copy_stream, pinned+B11.tileroff(j)*B11.tilecols(i-1));
+#if defined(STRUMPACK_USE_OPENMP_TASKLOOP)
 #pragma omp taskloop
+#endif
               for (std::size_t j=0; j<rb2; j++)
                 B12.tile(i-1, j).move_to_cpu
                   (copy_stream, pinned+B12.tilecoff(j)*B12.tilerows(i-1));
+#if defined(STRUMPACK_USE_OPENMP_TASKLOOP)
 #pragma omp taskloop
+#endif
               for (std::size_t j=0; j<rb2; j++)
                 B21.tile(j, i-1).move_to_cpu
                   (copy_stream, pinned+B21.tileroff(j)*B21.tilecols(i-1));

src/BLR/BLRMatrix.cpp

@@ -1401,7 +1401,7 @@ namespace strumpack {
 #endif
         {
           for (std::size_t i=0; i<rb; i+=CP) { // F11 and F21
-#pragma omp taskwait
+	    // #pragma omp taskwait
 #if defined(STRUMPACK_USE_OPENMP_TASK_DEPEND)
             [[maybe_unused]] std::size_t ifirst = lrb*i;
 #pragma omp task default(shared) firstprivate(i,ifirst) \
@@ -1534,9 +1534,10 @@ namespace strumpack {
                 }
               }
             }
-          }
-          for (std::size_t i=0; i<rb2; i+=CP) { // F12 and F22
 #pragma omp taskwait
+	  }
+          for (std::size_t i=0; i<rb2; i+=CP) { // F12 and F22
+	    // #pragma omp taskwait
 #if defined(STRUMPACK_USE_OPENMP_TASK_DEPEND)
             [[maybe_unused]] std::size_t ifirst = lrb*(i+rb);
 #pragma omp task default(shared) firstprivate(i,ifirst) \
@@ -1603,7 +1604,8 @@ namespace strumpack {
                 }
               }
             }
-          }
+#pragma omp taskwait
+	  }
         }
       }
       for (std::size_t i=0; i<rb; i++)

src/dense/DenseMatrix.hpp

@@ -1032,7 +1032,7 @@ namespace strumpack {
      * \param ld leading dimension of matrix allocated at D. ld >= m
      */
     DenseMatrixWrapper(std::size_t m, std::size_t n,
-                       scalar_t* D, std::size_t ld) {
+                       scalar_t* D, std::size_t ld) noexcept {
       this->data_ = D; this->rows_ = m; this->cols_ = n;
       this->ld_ = std::max(std::size_t(1), ld);
     }
@@ -1051,13 +1051,13 @@ namespace strumpack {
      * submatrix
      */
     DenseMatrixWrapper(std::size_t m, std::size_t n, DenseMatrix<scalar_t>& D,
-                       std::size_t i, std::size_t j)
+                       std::size_t i, std::size_t j) noexcept
       : DenseMatrixWrapper<scalar_t>(m, n, &D(i, j), D.ld()) {
       assert(i+m <= D.rows());
       assert(j+n <= D.cols());
     }
 
-    DenseMatrixWrapper(DenseMatrixWrapper<scalar_t>& D) {
+    DenseMatrixWrapper(DenseMatrixWrapper<scalar_t>& D) noexcept {
       this->data_ = D.data();
       this->rows_ = D.rows();
       this->cols_ = D.cols();
@@ -1067,7 +1067,7 @@ namespace strumpack {
     /**
      * Default move constructor.
      */
-    DenseMatrixWrapper(DenseMatrixWrapper<scalar_t>&& D) {
+    DenseMatrixWrapper(DenseMatrixWrapper<scalar_t>&& D) noexcept {
       this->data_ = D.data();
       this->rows_ = D.rows();
       this->cols_ = D.cols();

src/structured/StructuredMatrix.cpp

@@ -225,7 +225,9 @@ namespace strumpack {
                   std::vector<std::size_t> J(n);
                   std::iota(J.begin(), J.end(), c);
                   DenseMW_t Rrsub(n, Rr.cols(), Rr, c, 0);
+#if defined(STRUMPACK_USE_OPENMP_TASKLOOP)
 #pragma omp taskloop firstprivate(c) shared(Rrsub)
+#endif
                   for (int r=0; r<rows; r+=B) {
                     int m = std::min(B, rows - r);
                     DenseM_t Asub(m, n);
@@ -242,7 +244,9 @@ namespace strumpack {
                   std::vector<std::size_t> I(m);
                   std::iota(I.begin(), I.end(), r);
                   DenseMW_t Rcsub(m, Rc.cols(), Rc, r, 0);
+#if defined(STRUMPACK_USE_OPENMP_TASKLOOP)
 #pragma omp taskloop firstprivate(r) shared(Rcsub)
+#endif
                   for (int c=0; c<cols; c+=B) {
                     int n = std::min(B, cols - c);
                     DenseM_t Asub(m, n);