explicit casting of zero

kokkos · May 6, 2024 · aa8440c · aa8440c
1 parent 20f9fde
commit aa8440c
Show file tree

Hide file tree

Showing 3 changed files with 15 additions and 12 deletions.
diff --git a/sparse/impl/KokkosSparse_spmv_impl.hpp b/sparse/impl/KokkosSparse_spmv_impl.hpp
@@ -450,17 +450,18 @@ static void spmv_beta_transpose(const execution_space& exec,
                                 const AMatrix& A, const XVector& x,
                                 typename YVector::const_value_type& beta,
                                 const YVector& y) {
-  using ordinal_type = typename AMatrix::non_const_ordinal_type;
-  using size_type    = typename AMatrix::non_const_size_type;
+  using ordinal_type  = typename AMatrix::non_const_ordinal_type;
+  using size_type     = typename AMatrix::non_const_size_type;
+  using y_scalar_type = typename YVector::non_const_value_type;
 
   if (A.numRows() <= static_cast<ordinal_type>(0)) {
     return;
   }
 
   // We need to scale y first ("scaling" by zero just means filling
   // with zeros), since the functor works by atomic-adding into y.
-  if (0 == dobeta || 0 == beta) {
-    Kokkos::deep_copy(exec, y, 0);
+  if (0 == dobeta || y_scalar_type(0) == beta) {
+    Kokkos::deep_copy(exec, y, y_scalar_type(0));
   } else if (dobeta != 1) {
     KokkosBlas::scal(exec, y, beta, y);
   }
@@ -542,17 +543,18 @@ static void spmv_beta_transpose(const execution_space& exec,
                                 const AMatrix& A, const XVector& x,
                                 typename YVector::const_value_type& beta,
                                 const YVector& y) {
-  using ordinal_type = typename AMatrix::non_const_ordinal_type;
-  using size_type    = typename AMatrix::non_const_size_type;
+  using ordinal_type  = typename AMatrix::non_const_ordinal_type;
+  using size_type     = typename AMatrix::non_const_size_type;
+  using y_scalar_type = typename YVector::non_const_value_type;
 
   if (A.numRows() <= static_cast<ordinal_type>(0)) {
     return;
   }
 
   // We need to scale y first ("scaling" by zero just means filling
   // with zeros), since the functor works by atomic-adding into y.
-  if (0 == dobeta || 0 == beta) {
-    Kokkos::deep_copy(exec, y, 0);
+  if (0 == dobeta || y_scalar_type(0) == beta) {
+    Kokkos::deep_copy(exec, y, y_scalar_type(0));
   } else if (dobeta != 1) {
     KokkosBlas::scal(exec, y, beta, y);
   }

diff --git a/sparse/impl/KokkosSparse_spmv_impl_merge.hpp b/sparse/impl/KokkosSparse_spmv_impl_merge.hpp
@@ -309,8 +309,8 @@ struct SpmvMergeHierarchical {
     static_assert(XVector::rank == 1, "");
     static_assert(YVector::rank == 1, "");
 
-    if (0 == beta) {
-      Kokkos::deep_copy(space, y, 0);
+    if (y_value_type(0) == beta) {
+      Kokkos::deep_copy(space, y, y_value_type(0));
     } else {
       KokkosBlas::scal(space, y, beta, y);
     }

diff --git a/sparse/unit_test/Test_Sparse_spmv.hpp b/sparse/unit_test/Test_Sparse_spmv.hpp
@@ -119,6 +119,7 @@ void sequential_spmv(crsMat_t input_mat, x_vector_type x, y_vector_type y,
   using size_type_view_t = typename graph_t::row_map_type;
   using lno_view_t       = typename graph_t::entries_type;
   using scalar_view_t    = typename crsMat_t::values_type::non_const_type;
+  using y_scalar_t       = typename y_vector_type::non_const_value_type;
 
   using size_type = typename size_type_view_t::non_const_value_type;
   using lno_t     = typename lno_view_t::non_const_value_type;
@@ -149,8 +150,8 @@ void sequential_spmv(crsMat_t input_mat, x_vector_type x, y_vector_type y,
 
   // first, scale y by beta
   for (size_t i = 0; i < h_y.extent(0); i++) {
-    if (beta == 0) {
-      h_y(i) = 0;
+    if (beta == y_scalar_t(0)) {
+      h_y(i) = y_scalar_t(0);
     } else {
       h_y(i) *= beta;
     }