assorted bug fixes for the std::execution implementation in cudax (#…

…3721)
NVIDIA · Feb 6, 2025 · 0d6ab79 · 0d6ab79
1 parent 044cabf
commit 0d6ab79
Show file tree

Hide file tree

Showing 6 changed files with 47 additions and 21 deletions.
diff --git a/cudax/include/cuda/experimental/__async/sender/env.cuh b/cudax/include/cuda/experimental/__async/sender/env.cuh
@@ -76,6 +76,9 @@ struct _CCCL_TYPE_VISIBILITY_DEFAULT prop
   prop& operator=(const prop&) = delete;
 };
 
+template <class _Query, class _Value>
+prop(_Query, _Value) -> prop<_Query, _Value>;
+
 template <class... _Envs>
 struct _CCCL_TYPE_VISIBILITY_DEFAULT env
 {

diff --git a/cudax/include/cuda/experimental/__async/sender/meta.cuh b/cudax/include/cuda/experimental/__async/sender/meta.cuh
@@ -59,6 +59,8 @@ struct _IN_ALGORITHM;
 
 struct _WHAT;
 
+struct _WHY;
+
 struct _WITH_FUNCTION;
 
 struct _WITH_SENDER;

diff --git a/cudax/include/cuda/experimental/__async/sender/rcvr_with_env.cuh b/cudax/include/cuda/experimental/__async/sender/rcvr_with_env.cuh
@@ -134,6 +134,10 @@ struct __rcvr_with_env_t<_Rcvr*, _Env>
   _Rcvr* __rcvr_;
   _Env __env_;
 };
+
+template <class _Rcvr, class _Env>
+__rcvr_with_env_t(_Rcvr, _Env) -> __rcvr_with_env_t<_Rcvr, _Env>;
+
 } // namespace cuda::experimental::__async
 
 #include <cuda/experimental/__async/sender/epilogue.cuh>

diff --git a/cudax/include/cuda/experimental/__async/sender/sync_wait.cuh b/cudax/include/cuda/experimental/__async/sender/sync_wait.cuh
@@ -26,13 +26,15 @@
 // run_loop isn't supported on-device yet, so neither can sync_wait be.
 #if !defined(__CUDA_ARCH__)
 
+#  include <cuda/std/__type_traits/type_identity.h>
 #  include <cuda/std/optional>
 #  include <cuda/std/tuple>
 
 #  include <cuda/experimental/__async/sender/exception.cuh>
 #  include <cuda/experimental/__async/sender/meta.cuh>
 #  include <cuda/experimental/__async/sender/run_loop.cuh>
 #  include <cuda/experimental/__async/sender/utility.cuh>
+#  include <cuda/experimental/__async/sender/write_env.cuh>
 
 #  include <system_error>
 
@@ -115,26 +117,38 @@ private:
       }
     };
 
-    using __values_t = value_types_of_t<_Sndr, __rcvr_t, _CUDA_VSTD::tuple, _CUDA_VSTD::__type_self_t>;
+    using __completions_t = completion_signatures_of_t<_Sndr, __rcvr_t>;
+
+    struct __on_success
+    {
+      using type = __value_types<__completions_t, _CUDA_VSTD::tuple, _CUDA_VSTD::__type_self_t>;
+    };
+
+    using __on_error = _CUDA_VSTD::type_identity<_CUDA_VSTD::tuple<__completions_t>>;
+
+    using __values_t =
+      typename _CUDA_VSTD::_If<__is_completion_signatures<__completions_t>, __on_success, __on_error>::type;
 
     _CUDA_VSTD::optional<__values_t>* __values_;
     ::std::exception_ptr __eptr_;
     run_loop __loop_;
   };
 
-  struct __invalid_sync_wait
+  template <class _Type>
+  struct __always_false : _CUDA_VSTD::false_type
+  {};
+
+  template <class _Diagnostic>
+  struct __bad_sync_wait
   {
-    const __invalid_sync_wait& value() const
-    {
-      return *this;
-    }
+    static_assert(__always_false<_Diagnostic>(),
+                  "sync_wait cannot compute the completions of the sender passed to it.");
+    static __bad_sync_wait __result();
 
-    const __invalid_sync_wait& operator*() const
-    {
-      return *this;
-    }
+    const __bad_sync_wait& value() const;
+    const __bad_sync_wait& operator*() const;
 
-    int __i_;
+    int i{}; // so that structured bindings kinda work
   };
 
 public:
@@ -168,12 +182,11 @@ public:
   {
     using __rcvr_t      = typename __state_t<_Sndr>::__rcvr_t;
     using __values_t    = typename __state_t<_Sndr>::__values_t;
-    using __completions = completion_signatures_of_t<_Sndr, __rcvr_t>;
-    static_assert(__is_completion_signatures<__completions>);
+    using __completions = typename __state_t<_Sndr>::__completions_t;
 
     if constexpr (!__is_completion_signatures<__completions>)
     {
-      return __invalid_sync_wait{0};
+      return __bad_sync_wait<__completions>::__result();
     }
     else
     {
@@ -196,6 +209,12 @@ public:
       return __result; // uses NRVO to "return" the result
     }
   }
+
+  template <class _Sndr, class _Env>
+  auto operator()(_Sndr&& __sndr, _Env&& __env) const
+  {
+    return (*this)(__async::write_env(static_cast<_Sndr&&>(__sndr), static_cast<_Env&&>(__env)));
+  }
 };
 
 _CCCL_GLOBAL_CONSTANT sync_wait_t sync_wait{};

diff --git a/cudax/include/cuda/experimental/__async/sender/utility.cuh b/cudax/include/cuda/experimental/__async/sender/utility.cuh
@@ -21,6 +21,8 @@
 #  pragma system_header
 #endif // no system header
 
+#include <cuda/std/__tuple_dir/ignore.h>
+#include <cuda/std/__type_traits/decay.h>
 #include <cuda/std/__type_traits/is_same.h>
 #include <cuda/std/initializer_list>
 
@@ -34,11 +36,7 @@ namespace cuda::experimental::__async
 {
 _CCCL_GLOBAL_CONSTANT size_t __npos = static_cast<size_t>(-1);
 
-struct __ignore
-{
-  template <class... _As>
-  _CUDAX_API constexpr __ignore(_As&&...) noexcept {};
-};
+using __ignore _CCCL_NODEBUG_ALIAS = _CUDA_VSTD::__ignore_t; // NOLINT: misc-unused-using-decls
 
 using _CUDA_VSTD::__undefined; // NOLINT: misc-unused-using-decls
 
@@ -116,7 +114,7 @@ _CUDAX_API constexpr void __swap(_Ty& __left, _Ty& __right) noexcept
 }
 
 template <class _Ty>
-_CUDAX_API constexpr _Ty __decay_copy(_Ty&& __ty) noexcept(__nothrow_decay_copyable<_Ty>)
+_CUDAX_API constexpr _CUDA_VSTD::decay_t<_Ty> __decay_copy(_Ty&& __ty) noexcept(__nothrow_decay_copyable<_Ty>)
 {
   return static_cast<_Ty&&>(__ty);
 }

diff --git a/cudax/include/cuda/experimental/__async/sender/write_env.cuh b/cudax/include/cuda/experimental/__async/sender/write_env.cuh
@@ -50,7 +50,7 @@ private:
     connect_result_t<_Sndr, __rcvr_with_env_t<_Rcvr, _Env>*> __opstate_;
 
     _CUDAX_API explicit __opstate_t(_Sndr&& __sndr, _Env __env, _Rcvr __rcvr)
-        : __env_rcvr_(static_cast<_Env&&>(__env), static_cast<_Rcvr&&>(__rcvr))
+        : __env_rcvr_{static_cast<_Rcvr&&>(__rcvr), static_cast<_Env&&>(__env)}
         , __opstate_(__async::connect(static_cast<_Sndr&&>(__sndr), &__env_rcvr_))
     {}