improve document, fix auto usage in for loop

Co-authored-by: Thomas Grützmacher <thomas.gruetzmacher@kit.edu>

common/components/prefix_sum.hpp.inc

@@ -57,7 +57,7 @@ __forceinline__ __device__ void subwarp_prefix_sum(ValueType element,
     total_sum = element;
 #pragma unroll
     // hypercube prefix sum
-    for (auto step = 1; step < subwarp.size(); step *= 2) {
+    for (int step = 1; step < subwarp.size(); step *= 2) {
         auto neighbor = subwarp.shfl_xor(total_sum, step);
         total_sum += neighbor;
         prefix_sum += bool(subwarp.thread_rank() & step) ? neighbor : 0;

common/components/sorting.hpp.inc

@@ -70,7 +70,7 @@ struct bitonic_local {
                                                           bool reverse)
     {
         auto els_mid = els + (num_elements / 2);
-        for (auto i = 0; i < num_elements / 2; ++i) {
+        for (int i = 0; i < num_elements / 2; ++i) {
             bitonic_cas(els[i], els_mid[i], reverse);
         }
         half::merge(els, reverse);
@@ -131,7 +131,7 @@ struct bitonic_warp {
         auto tile =
             group::tiled_partition<num_threads>(group::this_thread_block());
         auto new_reverse = reverse != upper_half();
-        for (auto i = 0; i < num_local; ++i) {
+        for (int i = 0; i < num_local; ++i) {
             auto other = tile.shfl_xor(els[i], num_threads / 2);
             bitonic_cas(els[i], other, new_reverse);
         }
@@ -206,7 +206,7 @@ struct bitonic_global {
         auto upper_shared_els = shared_els + (num_groups * num_threads / 2);
         // only the lower group executes the CAS
         if (!upper_half()) {
-            for (auto i = 0; i < num_local; ++i) {
+            for (int i = 0; i < num_local; ++i) {
                 auto j = shared_idx(i);
                 bitonic_cas(shared_els[j], upper_shared_els[j], reverse);
             }
@@ -241,11 +241,11 @@ struct bitonic_global<ValueType, num_local, num_threads, 1, num_total_threads> {
                                                  bool reverse)
     {
         group::this_thread_block().sync();
-        for (auto i = 0; i < num_local; ++i) {
+        for (int i = 0; i < num_local; ++i) {
             local_els[i] = shared_els[shared_idx(i)];
         }
         warp::merge(local_els, reverse);
-        for (auto i = 0; i < num_local; ++i) {
+        for (int i = 0; i < num_local; ++i) {
             shared_els[shared_idx(i)] = local_els[i];
         }
     }
@@ -258,7 +258,7 @@ struct bitonic_global<ValueType, num_local, num_threads, 1, num_total_threads> {
         // This is the first step, so we don't need to load from shared memory
         warp::sort(local_els, reverse);
         // store the sorted elements in shared memory
-        for (auto i = 0; i < num_local; ++i) {
+        for (int i = 0; i < num_local; ++i) {
             shared_els[shared_idx(i)] = local_els[i];
         }
     }

common/components/uninitialized_array.hpp.inc

@@ -34,7 +34,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 /**
  * Stores an array with uninitialized contents.
  *
- * This class needed for datatypes that do have a non-empty constructor when`
+ * This class is needed for datatypes that do have a non-empty constructor when
  * using them as shared memory, for example `thrust::complex<float>`.
  *
  * @tparam ValueType the type of values
@@ -49,7 +49,7 @@ public:
      *
      * @return the constexpr pointer to the first entry of the array.
      */
-    constexpr GKO_ATTRIBUTES operator ValueType *() const noexcept
+    constexpr GKO_ATTRIBUTES operator const ValueType *() const noexcept
     {
         return &(*this)[0];
     }
@@ -70,7 +70,8 @@ public:
      *
      * @return a reference to the array entry at the given index.
      */
-    constexpr GKO_ATTRIBUTES ValueType &operator[](size_type pos) const noexcept
+    constexpr GKO_ATTRIBUTES const ValueType &operator[](size_type pos) const
+        noexcept
     {
         return reinterpret_cast<const ValueType *>(data_)[pos];
     }

common/factorization/par_ilut_filter_kernels.hpp.inc

@@ -55,7 +55,7 @@ __device__ void abstract_filter_impl(const IndexType *row_ptrs,
     auto end = row_ptrs[row + 1];
     begin_cb(row);
     auto num_steps = ceildiv(end - begin, subwarp_size);
-    for (auto step = 0; step < num_steps; ++step) {
+    for (IndexType step = 0; step < num_steps; ++step) {
         auto idx = begin + lane + step * subwarp_size;
         auto keep = idx < end && pred(idx, begin, end);
         auto mask = subwarp.ballot(keep);
@@ -189,4 +189,4 @@ __global__ __launch_bounds__(default_block_size) void bucket_filter(
 }
 
 
-}  // namespace kernel
+}  // namespace kernel