Add unique_ptr overloads for storeChunk calls (#1294)

* Add OpenpmdUniquePtr class

* Prepare IOTask.hpp for a non-copyable parameter

Refactor ::clone() to ::to_heap(), also make copy/move
constructors/operators in AbstractParameter protected instead of
deleting them.

* Backend fully prepared for accepting unique_ptr buffers

No optimizations based on that yet

* Add storeChunk(OpenpmdUniquePtr, ...) overload

* Fix invasive tests

* Add test for ADIOS2 backend optimization

Not yet implemented, so test fails

* Implement ADIOS2 backend optimization

* Support also regular std::unique_ptr

* CI fixes

* Use OpenpmdUniquePtr with storeChunk in an example

* Rename OpenpmdUniquePtr -> UniquePtrWithLambda

Author: franzpoeschel

examples/12_span_write.cpp

@@ -82,6 +82,36 @@ void span_write(std::string const &filename)
             }
             ++j;
         }
+
+        using mesh_type = position_t;
+
+        RecordComponent chargeDensity =
+            iteration.meshes["e_chargeDensity"][RecordComponent::SCALAR];
+
+        /*
+         * A similar memory optimization is possible by using a unique_ptr type
+         * in the call to storeChunk().
+         * Unlike the Span API, the buffer here is user-created, but in both
+         * approaches, the backend will manage the memory after the call to
+         * storeChunk().
+         * Some backends (especially: ADIOS2 BP5) will benefit from being able
+         * to avoid memcopies since they know that they can just keep the memory
+         * and noone else is reading it.
+         */
+        chargeDensity.resetDataset(dataset);
+        /*
+         * The class template UniquePtrWithLambda (subclass of std::unique_ptr)
+         * can be used to specify custom destructors, e.g. for deallocating
+         * GPU pointers.
+         * Normal std::unique_ptr types can also be used, even with custom
+         * destructors.
+         */
+        UniquePtrWithLambda<mesh_type> data(
+            new mesh_type[length](), [](auto const *ptr) { delete[] ptr; });
+        /*
+         * Move the unique_ptr into openPMD. It must now no longer be accessed.
+         */
+        chargeDensity.storeChunk(std::move(data), {0}, extent);
         iteration.close();
     }
 

include/openPMD/Datatype.hpp

@@ -21,6 +21,7 @@
 #pragma once
 
 #include "openPMD/auxiliary/TypeTraits.hpp"
+#include "openPMD/auxiliary/UniquePtr.hpp"
 
 #include <array>
 #include <climits>
@@ -276,31 +277,26 @@ inline constexpr Datatype determineDatatype()
         return Datatype::UNDEFINED;
 }
 
-template <typename T>
-inline constexpr Datatype determineDatatype(std::shared_ptr<T>)
-{
-    return determineDatatype<T>();
-}
-
-template <typename T>
-inline constexpr Datatype determineDatatype(T *)
-{
-    return determineDatatype<T>();
-}
-
-/*
- * Catch-all overload for unsupported types, with static_assert errors
- * triggered at compile-time.
+/**
+ * @brief Determine datatype of passed value
+ *
+ * @param val Value whose type to evaluate
+ * @tparam T Type of the passed value
+ * @return If T is of a pointer type, then the type of the contained value.
+ *         Otherwise, a compile-time error detailing the use of this function.
  */
-template <typename T_ContiguousContainer>
-inline constexpr Datatype determineDatatype(T_ContiguousContainer &&)
+template <typename T>
+inline constexpr Datatype determineDatatype(T &&val)
 {
-    using T_ContiguousContainer_stripped =
-        std::remove_reference_t<T_ContiguousContainer>;
-    if constexpr (auxiliary::IsContiguousContainer_v<
-                      T_ContiguousContainer_stripped>)
+    (void)val; // don't need this, it only has a name for Doxygen
+    using T_stripped = std::remove_cv_t<std::remove_reference_t<T>>;
+    if constexpr (auxiliary::IsPointer_v<T_stripped>)
+    {
+        return determineDatatype<auxiliary::IsPointer_t<T_stripped>>();
+    }
+    else if constexpr (auxiliary::IsContiguousContainer_v<T_stripped>)
     {
-        static_assert(auxiliary::dependent_false_v<T_ContiguousContainer>, R"(
+        static_assert(auxiliary::dependent_false_v<T_stripped>, R"(
 Error: Passed a contiguous container type to determineDatatype<>().
 These types are not directly supported due to colliding semantics.
 Assuming a vector object `std::vector<float> vec;`,
@@ -322,15 +318,14 @@ use one of the following alternatives:
     }
     else
     {
-        static_assert(auxiliary::dependent_false_v<T_ContiguousContainer>, R"(
+        static_assert(auxiliary::dependent_false_v<T_stripped>, R"(
 Error: Unknown datatype passed to determineDatatype<>().
 For a direct translation from C++ type to the openPMD::Datatype enum, use:
 `auto determineDatatype<T>() -> Datatype`.
 
 For detecting the contained datatpye of a pointer type (shared or raw pointer),
-use either of the following overloads:
-`auto determineDatatype<T>(std::shared_ptr<T>) -> Datatype` or
-`auto determineDatatype<T>(T *) -> Datatype`.
+use this following template (i.e. `auto determineDatatype<T>(T &&) -> Datatype`)
+which accepts pointer-type parameters (raw, shared or unique).
         )");
     }
     // Unreachable, but C++ does not know it

include/openPMD/IO/ADIOS/ADIOS2IOHandler.hpp

@@ -82,6 +82,7 @@ namespace detail
     struct BufferedGet;
     struct BufferedAttributeRead;
     struct BufferedAttributeWrite;
+    struct RunUniquePtrPut;
 } // namespace detail
 
 namespace ADIOS2Schema
@@ -124,6 +125,7 @@ class ADIOS2IOHandlerImpl
     friend struct detail::WriteDataset;
     friend struct detail::BufferedActions;
     friend struct detail::BufferedAttributeRead;
+    friend struct detail::RunUniquePtrPut;
 
 public:
 #if openPMD_HAVE_MPI
@@ -188,8 +190,8 @@ class ADIOS2IOHandlerImpl
     void deleteAttribute(
         Writable *, Parameter<Operation::DELETE_ATT> const &) override;
 
-    void writeDataset(
-        Writable *, Parameter<Operation::WRITE_DATASET> const &) override;
+    void
+    writeDataset(Writable *, Parameter<Operation::WRITE_DATASET> &) override;
 
     void writeAttribute(
         Writable *, Parameter<Operation::WRITE_ATT> const &) override;
@@ -544,11 +546,7 @@ namespace detail
     struct WriteDataset
     {
         template <typename T>
-        static void call(
-            ADIOS2IOHandlerImpl *impl,
-            BufferedPut &bp,
-            adios2::IO &IO,
-            adios2::Engine &engine);
+        static void call(BufferedActions &ba, BufferedPut &bp);
 
         template <int n, typename... Params>
         static void call(Params &&...);
@@ -916,6 +914,17 @@ namespace detail
         void run(BufferedActions &) override;
     };
 
+    struct BufferedUniquePtrPut
+    {
+        std::string name;
+        Offset offset;
+        Extent extent;
+        UniquePtrWithLambda<void> data;
+        Datatype dtype;
+
+        void run(BufferedActions &);
+    };
+
     struct OldBufferedAttributeRead : BufferedAction
     {
         Parameter<Operation::READ_ATT> param;
@@ -967,6 +976,8 @@ namespace detail
     {
         friend struct BufferedGet;
         friend struct BufferedPut;
+        friend struct RunUniquePtrPut;
+        friend struct WriteDataset;
 
         using FlushTarget = ADIOS2IOHandlerImpl::FlushTarget;
 
@@ -1023,6 +1034,13 @@ namespace detail
          * penalty, once preloadAttributes has been filled.
          */
         std::vector<BufferedAttributeRead> m_attributeReads;
+        /**
+         * When receiving a unique_ptr, we know that the buffer is ours and
+         * ours alone. So, for performance reasons, show the buffer to ADIOS2 as
+         * late as possible and avoid unnecessary data copies in BP5 triggered
+         * by PerformDataWrites().
+         */
+        std::vector<BufferedUniquePtrPut> m_uniquePtrPuts;
         /**
          * This contains deferred actions that have already been enqueued into
          * ADIOS2, but not yet performed in ADIOS2.
@@ -1116,24 +1134,26 @@ namespace detail
          *     * adios2::Engine::EndStep
          *     * adios2::Engine::Perform(Puts|Gets)
          *     * adios2::Engine::Close
-         * @param writeAttributes If using the new attribute layout, perform
-         *     deferred attribute writes now.
+         * @param writeLatePuts Some things are deferred until right before
+         *        Engine::EndStep() or Engine::Close():
+         *        1) Writing attributes in new ADIOS2 schema.
+         *        2) Running unique_ptr Put()s.
          * @param flushUnconditionally Whether to run the functor even if no
          *     deferred IO tasks had been queued.
          */
         template <typename F>
         void flush_impl(
             ADIOS2FlushParams flushParams,
             F &&performPutsGets,
-            bool writeAttributes,
+            bool writeLatePuts,
             bool flushUnconditionally);
 
         /**
          * Overload of flush() that uses adios2::Engine::Perform(Puts|Gets)
          * and does not flush unconditionally.
          *
          */
-        void flush_impl(ADIOS2FlushParams, bool writeAttributes = false);
+        void flush_impl(ADIOS2FlushParams, bool writeLatePuts = false);
 
         /**
          * @brief Begin or end an ADIOS step.

include/openPMD/IO/ADIOS/CommonADIOS1IOHandler.hpp

@@ -74,8 +74,8 @@ class CommonADIOS1IOHandlerImpl : public AbstractIOHandlerImpl
         Writable *, Parameter<Operation::DELETE_DATASET> const &) override;
     void deleteAttribute(
         Writable *, Parameter<Operation::DELETE_ATT> const &) override;
-    void writeDataset(
-        Writable *, Parameter<Operation::WRITE_DATASET> const &) override;
+    void
+    writeDataset(Writable *, Parameter<Operation::WRITE_DATASET> &) override;
     void writeAttribute(
         Writable *, Parameter<Operation::WRITE_ATT> const &) override;
     void readDataset(Writable *, Parameter<Operation::READ_DATASET> &) override;

include/openPMD/IO/AbstractIOHandlerImpl.hpp

@@ -456,7 +456,7 @@ class AbstractIOHandlerImpl
      * storage after the operation completes successfully.
      */
     virtual void
-    writeDataset(Writable *, Parameter<Operation::WRITE_DATASET> const &) = 0;
+    writeDataset(Writable *, Parameter<Operation::WRITE_DATASET> &) = 0;
 
     /** Get a view into a dataset buffer that can be filled by a user.
      *

include/openPMD/IO/HDF5/HDF5IOHandlerImpl.hpp

@@ -65,8 +65,8 @@ class HDF5IOHandlerImpl : public AbstractIOHandlerImpl
         Writable *, Parameter<Operation::DELETE_DATASET> const &) override;
     void deleteAttribute(
         Writable *, Parameter<Operation::DELETE_ATT> const &) override;
-    void writeDataset(
-        Writable *, Parameter<Operation::WRITE_DATASET> const &) override;
+    void
+    writeDataset(Writable *, Parameter<Operation::WRITE_DATASET> &) override;
     void writeAttribute(
         Writable *, Parameter<Operation::WRITE_ATT> const &) override;
     void readDataset(Writable *, Parameter<Operation::READ_DATASET> &) override;