Add scoped_trace and fix omp kernel.

ginkgo-project · Feb 27, 2022 · 089cdc0 · 089cdc0
1 parent acd5d7c
commit 089cdc0
Show file tree

Hide file tree

Showing 2 changed files with 12 additions and 4 deletions.
diff --git a/omp/matrix/csr_kernels.cpp b/omp/matrix/csr_kernels.cpp
@@ -755,12 +755,11 @@ void calculate_nonzeros_per_row_in_index_set(
     auto row_subset_begin = row_index_set.get_subsets_begin();
     auto row_subset_end = row_index_set.get_subsets_end();
     auto src_ptrs = source->get_const_row_ptrs();
-#pragma omp parallel for
+    size_type res_row = 0;
     for (size_type set = 0; set < num_row_subsets; ++set) {
         for (size_type row = row_subset_begin[set]; row < row_subset_end[set];
              ++row) {
-            row_nnz->get_data()[row - row_subset_begin[set]] =
-                zero<IndexType>();
+            row_nnz->get_data()[res_row] = zero<IndexType>();
             Array<IndexType> l_idxs(
                 exec,
                 static_cast<size_type>(src_ptrs[row + 1] - src_ptrs[row]));
@@ -776,9 +775,10 @@ void calculate_nonzeros_per_row_in_index_set(
                  ++nnz) {
                 auto l_idx = l_idxs.get_const_data()[nnz];
                 if (l_idx != invalid_index<IndexType>()) {
-                    row_nnz->get_data()[row - row_subset_begin[set]]++;
+                    row_nnz->get_data()[res_row]++;
                 }
             }
+            res_row++;
         }
     }
 }

diff --git a/reference/test/matrix/csr_kernels.cpp b/reference/test/matrix/csr_kernels.cpp
@@ -1699,27 +1699,31 @@ TYPED_TEST(Csr, CanGetSubmatrix2)
         this->exec);
     ASSERT_EQ(mat->get_num_stored_elements(), 23);
     {
+        SCOPED_TRACE("Left top corner: Square 2x2");
         auto sub_mat1 = mat->create_submatrix(gko::span(0, 2), gko::span(0, 2));
         auto ref1 =
             gko::initialize<Mtx>({I<T>{1.0, 3.0}, I<T>{1.0, 0.0}}, this->exec);
 
         GKO_EXPECT_MTX_NEAR(sub_mat1.get(), ref1.get(), 0.0);
     }
     {
+        SCOPED_TRACE("Left boundary: Square 2x2");
         auto sub_mat2 = mat->create_submatrix(gko::span(2, 4), gko::span(0, 2));
         auto ref2 =
             gko::initialize<Mtx>({I<T>{0.0, 3.0}, I<T>{0.0, -1.0}}, this->exec);
 
         GKO_EXPECT_MTX_NEAR(sub_mat2.get(), ref2.get(), 0.0);
     }
     {
+        SCOPED_TRACE("Right boundary: Square 2x2");
         auto sub_mat3 = mat->create_submatrix(gko::span(0, 2), gko::span(3, 5));
         auto ref3 =
             gko::initialize<Mtx>({I<T>{0.0, 2.0}, I<T>{7.5, 3.0}}, this->exec);
 
         GKO_EXPECT_MTX_NEAR(sub_mat3.get(), ref3.get(), 0.0);
     }
     {
+        SCOPED_TRACE("Non-square 5x2");
         auto sub_mat4 = mat->create_submatrix(gko::span(1, 6), gko::span(2, 4));
         /*
            4.5, 7.5
@@ -1779,6 +1783,7 @@ TYPED_TEST(Csr, CanGetSubmatrixWithIndexSet)
         this->exec);
     ASSERT_EQ(mat->get_num_stored_elements(), 23);
     {
+        SCOPED_TRACE("Small square 2x2");
         auto row_set = gko::IndexSet<index_type>(this->exec, {0, 1});
         auto col_set = gko::IndexSet<index_type>(this->exec, {0, 1});
         auto sub_mat1 = mat->create_submatrix(row_set, col_set);
@@ -1789,6 +1794,7 @@ TYPED_TEST(Csr, CanGetSubmatrixWithIndexSet)
     }
 
     {
+        SCOPED_TRACE("Non-square 4x2");
         auto row_set = gko::IndexSet<index_type>(this->exec, {1, 2, 3, 4});
         auto col_set = gko::IndexSet<index_type>(this->exec, {1, 3});
         auto sub_mat1 = mat->create_submatrix(row_set, col_set);
@@ -1800,6 +1806,7 @@ TYPED_TEST(Csr, CanGetSubmatrixWithIndexSet)
     }
 
     {
+        SCOPED_TRACE("Square 3x3");
         auto row_set = gko::IndexSet<index_type>(this->exec, {1, 3, 4});
         auto col_set = gko::IndexSet<index_type>(this->exec, {1, 3, 0});
         auto sub_mat1 = mat->create_submatrix(row_set, col_set);
@@ -1811,6 +1818,7 @@ TYPED_TEST(Csr, CanGetSubmatrixWithIndexSet)
     }
 
     {
+        SCOPED_TRACE("Square 4x4");
         auto row_set = gko::IndexSet<index_type>(this->exec, {1, 4, 5, 6});
         auto col_set = gko::IndexSet<index_type>(this->exec, {4, 3, 0, 1});
         auto sub_mat1 = mat->create_submatrix(row_set, col_set);