Merge pull request #347 from crtrott/cuda-fixes

More CUDA fixes exposed during Kokkos integration
kokkos · Jun 17, 2024 · 6f20467 · 6f20467
2 parents e01c4c6 + 99d810b
commit 6f20467
Show file tree

Hide file tree

Showing 3 changed files with 26 additions and 4 deletions.
diff --git a/include/experimental/__p0009_bits/extents.hpp b/include/experimental/__p0009_bits/extents.hpp
@@ -33,6 +33,7 @@ namespace detail {
 // Function used to check compatibility of extents in converting constructor
 // can't be a private member function for some reason.
 template <size_t... Extents, size_t... OtherExtents>
+MDSPAN_INLINE_FUNCTION
 static constexpr std::integral_constant<bool, false> __check_compatible_extents(
     std::integral_constant<bool, false>,
     std::integer_sequence<size_t, Extents...>,
@@ -49,6 +50,7 @@ struct __compare_extent_compatible : std::integral_constant<bool,
 {};
 
 template <size_t... Extents, size_t... OtherExtents>
+MDSPAN_INLINE_FUNCTION
 static constexpr std::integral_constant<
     bool, _MDSPAN_FOLD_AND(__compare_extent_compatible<Extents, OtherExtents>::value)>
 __check_compatible_extents(

diff --git a/include/experimental/__p0009_bits/utility.hpp b/include/experimental/__p0009_bits/utility.hpp
@@ -14,6 +14,7 @@ template <std::size_t N>
 using with_rank = std::integral_constant<std::size_t, N>;
 
 template <class I1, class I2>
+MDSPAN_INLINE_FUNCTION
 constexpr bool common_integral_compare(I1 x, I2 y)
 {
   static_assert(std::is_integral<I1>::value &&
@@ -24,11 +25,14 @@ constexpr bool common_integral_compare(I1 x, I2 y)
 }
 
 template <class T1, class T2, class F>
+MDSPAN_INLINE_FUNCTION
 constexpr bool rankwise_equal(with_rank<0>, const T1&, const T2&, F)
 {
   return true;
 }
+
 template <std::size_t N, class T1, class T2, class F>
+MDSPAN_INLINE_FUNCTION
 constexpr bool rankwise_equal(with_rank<N>, const T1& x, const T2& y, F func)
 {
   bool match = true;
@@ -43,6 +47,7 @@ constexpr bool rankwise_equal(with_rank<N>, const T1& x, const T2& y, F func)
 constexpr struct
 {
   template <class T, class I>
+  MDSPAN_INLINE_FUNCTION
   constexpr auto operator()(const T& x, I i) const
   {
     return x.extent(i);
@@ -52,6 +57,7 @@ constexpr struct
 constexpr struct
 {
   template <class T, class I>
+  MDSPAN_INLINE_FUNCTION
   constexpr auto operator()(const T& x, I i) const
   {
     return x.stride(i);

diff --git a/include/experimental/__p2630_bits/submdspan_mapping.hpp b/include/experimental/__p2630_bits/submdspan_mapping.hpp
@@ -284,8 +284,19 @@ MDSPAN_IMPL_PROPOSED_NAMESPACE::layout_left_padded<PaddingValue>::mapping<Extent
         out_of_bounds ? this->required_span_size()
                     : this->operator()(MDSPAN_IMPL_STANDARD_NAMESPACE::detail::first_of(slices)...));
     if constexpr (dst_ext_t::rank() == 0) { // result rank-0
-      using dst_mapping_t = typename layout_left::template mapping<dst_ext_t>;
-      return submdspan_mapping_result<dst_mapping_t>{dst_mapping_t{dst_ext}, offset};
+      // The following for some reasons leads to compiler error later, while not using a typedef works:
+      // Compilers: CUDA 11.2 with GCC 9.1
+      //
+      // using dst_mapping_t = typename layout_left::template mapping<dst_ext_t>;
+      // return submdspan_mapping_result<dst_mapping_t>{dst_mapping_t{dst_ext}, offset};
+      //
+      // Error: submdspan_mapping.hpp:299:23: error: 'dst_mapping_t' does not name a type
+      //         299 |         using dst_mapping_t = typename layout_left::template mapping<dst_ext_t>;
+      // The same error is given (about dst_mapping_t not naming type) when a different name is used in 299:
+      //        using dst_mapping_t2 = typename layout_left::template mapping<dst_ext_t>;
+
+      return submdspan_mapping_result<typename layout_left::template mapping<dst_ext_t>>
+             {typename layout_left::template mapping<dst_ext_t>{dst_ext}, offset};
     } else { // general case
       // Figure out if any slice's lower bound equals the corresponding extent.
       // If so, bypass evaluating the layout mapping.  This fixes LWG Issue 4060.
@@ -509,8 +520,11 @@ MDSPAN_IMPL_PROPOSED_NAMESPACE::layout_right_padded<PaddingValue>::mapping<Exten
         out_of_bounds ? this->required_span_size()
                     : this->operator()(MDSPAN_IMPL_STANDARD_NAMESPACE::detail::first_of(slices)...));
     if constexpr (dst_ext_t::rank() == 0) { // result rank-0
-      using dst_mapping_t = typename layout_right::template mapping<dst_ext_t>;
-      return submdspan_mapping_result<dst_mapping_t>{dst_mapping_t{dst_ext}, offset};
+      // Same issue as in layout_left_padded: see comment there
+      // using dst_mapping_t = typename layout_right::template mapping<dst_ext_t>;
+      // return submdspan_mapping_result<dst_mapping_t>{dst_mapping_t{dst_ext}, offset};
+      return submdspan_mapping_result<typename layout_right::template mapping<dst_ext_t>>
+        {typename layout_right::template mapping<dst_ext_t>{dst_ext}, offset};
     } else { // general case
       using deduce_layout = MDSPAN_IMPL_STANDARD_NAMESPACE::detail::deduce_layout_right_submapping<
         typename dst_ext_t::index_type, dst_ext_t::rank(),