SYCL: fixing issues revealed with icpx 2025 (kokkos#2669)

* SYCL: fixing issues revealed with icpx 2025

Removing some volatile attributes in spgemm and fixing issue with
reduction update variable initialization which would cause access
out of bound and segfaults.

Signed-off-by: Luc Berger-Vergiat <[email protected]>

* Fixing test and applying clang-format

Signed-off-by: Luc Berger-Vergiat <[email protected]>

---------

Signed-off-by: Luc Berger-Vergiat <[email protected]>

Author: lucbv

blas/impl/KokkosBlas1_iamax_impl.hpp

@@ -61,9 +61,7 @@ struct V_Iamax_Functor {
     if (val > maxval) lmaxloc = i;
   }
 
-  KOKKOS_INLINE_FUNCTION void init(value_type& update) const {
-    update = Kokkos::reduction_identity<typename RV::value_type>::max() + 1;
-  }
+  KOKKOS_INLINE_FUNCTION void init(value_type& update) const { update = 1; }
 
   KOKKOS_INLINE_FUNCTION void join(value_type& update, const value_type& source) const {
     mag_type source_val = IPT::norm(m_x(source - 1));

blas/unit_test/Test_Blas1_iamax.hpp

@@ -13,6 +13,7 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //@HEADER
+
 #include <gtest/gtest.h>
 #include <Kokkos_Core.hpp>
 #include <Kokkos_Random.hpp>
@@ -22,9 +23,9 @@
 namespace Test {
 template <class ViewTypeA, class Device>
 void impl_test_iamax(int N) {
-  typedef typename ViewTypeA::non_const_value_type ScalarA;
-  typedef Kokkos::ArithTraits<ScalarA> AT;
-  typedef typename AT::mag_type mag_type;
+  using ScalarA   = typename ViewTypeA::non_const_value_type;
+  using AT        = Kokkos::ArithTraits<ScalarA>;
+  using mag_type  = typename AT::mag_type;
   using size_type = typename ViewTypeA::size_type;
 
   view_stride_adapter<ViewTypeA> a("X", N);

sparse/impl/KokkosSparse_spgemm_impl_compression.hpp

@@ -573,14 +573,11 @@ struct KokkosSPGEMM<HandleType, a_row_view_t_, a_lno_nnz_view_t_, a_scalar_nnz_v
       if (overall_num_unsuccess) {
         // then we allocate second level memory using memory pool.
         if (!l2_allocated) {
-          volatile nnz_lno_t *tmp = NULL;
+          nnz_lno_t *tmp = NULL;
           while (tmp == NULL) {
             Kokkos::single(
                 Kokkos::PerThread(teamMember),
-                [&](volatile nnz_lno_t *&memptr) {
-                  memptr = (volatile nnz_lno_t *)(memory_space.allocate_chunk(row_ind));
-                },
-                tmp);
+                [&](nnz_lno_t *&memptr) { memptr = (nnz_lno_t *)(memory_space.allocate_chunk(row_ind)); }, tmp);
           }
           globally_used_hash_indices = (nnz_lno_t *)tmp;
           hm2.hash_begins            = (nnz_lno_t *)(globally_used_hash_indices + pow2_hash_size);

sparse/impl/KokkosSparse_spgemm_impl_symbolic.hpp

@@ -455,13 +455,12 @@ struct KokkosSPGEMM<HandleType, a_row_view_t_, a_lno_nnz_view_t_, a_scalar_nnz_v
         if (overall_num_unsuccess) {
           // printf("row:%d\n", row_index);
           if (!is_global_alloced) {
-            volatile nnz_lno_t *tmp = NULL;
-            size_t tid              = get_thread_id(row_index);
+            nnz_lno_t *tmp = NULL;
+            size_t tid     = get_thread_id(row_index);
             while (tmp == NULL) {
               Kokkos::single(
                   Kokkos::PerThread(teamMember),
-                  [&](volatile nnz_lno_t *&memptr) { memptr = (volatile nnz_lno_t *)(m_space.allocate_chunk(tid)); },
-                  tmp);
+                  [&](nnz_lno_t *&memptr) { memptr = (nnz_lno_t *)(m_space.allocate_chunk(tid)); }, tmp);
             }
             is_global_alloced = true;
 
@@ -1000,13 +999,12 @@ struct KokkosSPGEMM<HandleType, a_row_view_t_, a_lno_nnz_view_t_, a_scalar_nnz_v
         if (overall_num_unsuccess) {
           // printf("row:%d\n", row_index);
           if (!is_global_alloced) {
-            volatile nnz_lno_t *tmp = NULL;
-            size_t tid              = get_thread_id(row_index);
+            nnz_lno_t *tmp = NULL;
+            size_t tid     = get_thread_id(row_index);
             while (tmp == NULL) {
               Kokkos::single(
                   Kokkos::PerThread(teamMember),
-                  [&](volatile nnz_lno_t *&memptr) { memptr = (volatile nnz_lno_t *)(m_space.allocate_chunk(tid)); },
-                  tmp);
+                  [&](nnz_lno_t *&memptr) { memptr = (nnz_lno_t *)(m_space.allocate_chunk(tid)); }, tmp);
             }
             is_global_alloced = true;