[OFFLOAD][OPENMP] 6.0 compatible interop interface (llvm#143491)

The following patch introduces a new interop interface implementation
with the following characteristics:

* It supports the new 6.0 prefer_type specification
* It supports both explicit objects (from interop constructs) and
implicit objects (from variant calls).
* Implements a per-thread reuse mechanism for implicit objects to reduce
overheads.
* It provides a plugin interface that allows selecting the supported
interop types, and managing all the backend related interop operations
(init, sync, ...).
* It enables cooperation with the OpenMP runtime to allow progress on
OpenMP synchronizations.
* It cleanups some vendor/fr_id mismatchs from the current query
routines.
* It supports extension to define interop callbacks for library cleanup.

Author: adurang

offload/include/OpenMP/InteropAPI.h

@@ -13,17 +13,69 @@
 
 #include "omp.h"
 
+#include "PerThreadTable.h"
 #include "omptarget.h"
 
 extern "C" {
 
 typedef enum kmp_interop_type_t {
   kmp_interop_type_unknown = -1,
-  kmp_interop_type_platform,
-  kmp_interop_type_device,
-  kmp_interop_type_tasksync,
+  kmp_interop_type_target,
+  kmp_interop_type_targetsync,
 } kmp_interop_type_t;
 
+struct interop_attrs_t {
+  bool inorder : 1;
+  int reserved : 31;
+
+  /// Check if the supported attributes are compatible with the current
+  ///   attributes. Only if an attribute is supported can the value be true,
+  ///   otherwise it needs to be false
+  bool checkSupportedOnly(interop_attrs_t supported) const {
+    return supported.inorder || (!supported.inorder && !inorder);
+  }
+};
+
+struct interop_spec_t {
+  int32_t fr_id;
+  interop_attrs_t attrs; // Common attributes
+  int64_t impl_attrs; // Implementation specific attributes (recognized by each
+                      // plugin)
+};
+
+struct interop_flags_t {
+  bool implicit : 1; // dispatch (true) or interop (false)
+  bool nowait : 1;   // has nowait flag
+  int reserved : 30;
+};
+
+struct interop_ctx_t {
+  uint32_t version; // version of the interface (current is 0)
+  interop_flags_t flags;
+  int gtid;
+};
+
+struct dep_pack_t {
+  int32_t ndeps;
+  int32_t ndeps_noalias;
+  kmp_depend_info_t *deplist;
+  kmp_depend_info_t *noalias_deplist;
+};
+
+struct omp_interop_val_t;
+
+typedef void ompx_interop_cb_t(omp_interop_val_t *interop, void *data);
+
+struct omp_interop_cb_instance_t {
+  ompx_interop_cb_t *cb;
+  void *data;
+
+  omp_interop_cb_instance_t(ompx_interop_cb_t *cb, void *data)
+      : cb(cb), data(data) {}
+
+  void operator()(omp_interop_val_t *interop) { cb(interop, data); }
+};
+
 /// The interop value type, aka. the interop object.
 typedef struct omp_interop_val_t {
   /// Device and interop-type are determined at construction time and fix.
@@ -34,10 +86,98 @@ typedef struct omp_interop_val_t {
   __tgt_device_info device_info;
   const kmp_interop_type_t interop_type;
   const intptr_t device_id;
-  const omp_foreign_runtime_ids_t vendor_id = cuda;
-  const intptr_t backend_type_id = omp_interop_backend_type_cuda_1;
+  omp_vendor_id_t vendor_id = omp_vendor_llvm;
+  tgt_foreign_runtime_id_t fr_id = tgt_fr_none;
+  interop_attrs_t attrs{false, 0}; // Common prefer specification attributes
+  int64_t impl_attrs = 0; // Implementation prefer specification attributes
+
+  // Constants
+  static constexpr int no_owner = -1; // This interop has no current owner
+
+  void *rtl_property = nullptr; // Plugin dependent information
+  // For implicitly created Interop objects (e.g., from a dispatch construct)
+  // who owns the object
+  int owner_gtid = no_owner;
+  // Marks whether the object was requested since the last time it was synced
+  bool clean = true;
+
+  typedef llvm::SmallVector<omp_interop_cb_instance_t> callback_list_t;
+
+  callback_list_t completion_cbs;
+
+  void reset() {
+    owner_gtid = no_owner;
+    markClean();
+    clearCompletionCbs();
+  }
+
+  llvm::Expected<DeviceTy &> getDevice() const;
+
+  bool hasOwner() const { return owner_gtid != no_owner; }
+
+  void setOwner(int gtid) { owner_gtid = gtid; }
+  bool isOwnedBy(int gtid) { return owner_gtid == gtid; }
+  bool isCompatibleWith(int32_t InteropType, const interop_spec_t &Spec);
+  bool isCompatibleWith(int32_t InteropType, const interop_spec_t &Spec,
+                        int64_t DeviceNum, int gtid);
+  void markClean() { clean = true; }
+  void markDirty() { clean = false; }
+  bool isClean() const { return clean; }
+
+  int32_t flush(DeviceTy &Device);
+  int32_t sync_barrier(DeviceTy &Device);
+  int32_t async_barrier(DeviceTy &Device);
+  int32_t release(DeviceTy &Device);
+
+  void addCompletionCb(ompx_interop_cb_t *cb, void *data) {
+    completion_cbs.push_back(omp_interop_cb_instance_t(cb, data));
+  }
+
+  int numCompletionCbs() const { return completion_cbs.size(); }
+  void clearCompletionCbs() { completion_cbs.clear(); }
+
+  void runCompletionCbs() {
+    for (auto &cbInstance : completion_cbs)
+      cbInstance(this);
+    clearCompletionCbs();
+  }
 } omp_interop_val_t;
 
 } // extern "C"
 
+struct InteropTableEntry {
+  using ContainerTy = typename std::vector<omp_interop_val_t *>;
+  using iterator = typename ContainerTy::iterator;
+
+  ContainerTy Interops;
+
+  static constexpr int reservedEntriesPerThread =
+      20; // reserve some entries to avoid reallocation
+
+  void add(omp_interop_val_t *obj) {
+    if (Interops.capacity() == 0)
+      Interops.reserve(reservedEntriesPerThread);
+    Interops.push_back(obj);
+  }
+
+  template <class ClearFuncTy> void clear(ClearFuncTy f) {
+    for (auto &Obj : Interops) {
+      f(Obj);
+    }
+  }
+
+  /// vector interface
+  int size() const { return Interops.size(); }
+  iterator begin() { return Interops.begin(); }
+  iterator end() { return Interops.end(); }
+  iterator erase(iterator it) { return Interops.erase(it); }
+};
+
+struct InteropTblTy
+    : public PerThreadTable<InteropTableEntry, omp_interop_val_t *> {
+  void clear();
+};
+
+void syncImplicitInterops(int gtid, void *event);
+
 #endif // OMPTARGET_OPENMP_INTEROP_API_H

offload/include/OpenMP/omp.h

@@ -80,15 +80,18 @@ typedef enum omp_interop_rc {
   omp_irc_other = -6
 } omp_interop_rc_t;
 
-typedef enum omp_interop_fr {
-  omp_ifr_cuda = 1,
-  omp_ifr_cuda_driver = 2,
-  omp_ifr_opencl = 3,
-  omp_ifr_sycl = 4,
-  omp_ifr_hip = 5,
-  omp_ifr_level_zero = 6,
-  omp_ifr_last = 7
-} omp_interop_fr_t;
+/* Foreign runtime values from OpenMP Additional Definitions document v2.1 */
+typedef enum tgt_foreign_runtime_id_t {
+  tgt_fr_none = 0,
+  tgt_fr_cuda = 1,
+  tgt_fr_cuda_driver = 2,
+  tgt_fr_opencl = 3,
+  tgt_fr_sycl = 4,
+  tgt_fr_hip = 5,
+  tgt_fr_level_zero = 6,
+  tgt_fr_hsa = 7,
+  tgt_fr_last = 8
+} tgt_foreign_runtime_id_t;
 
 typedef void *omp_interop_t;
 
@@ -134,19 +137,23 @@ omp_get_interop_type_desc(const omp_interop_t, omp_interop_property_t);
 extern const char *__KAI_KMPC_CONVENTION
 omp_get_interop_rc_desc(const omp_interop_t, omp_interop_rc_t);
 
-typedef enum omp_interop_backend_type_t {
-  // reserve 0
-  omp_interop_backend_type_cuda_1 = 1,
-} omp_interop_backend_type_t;
-
-typedef enum omp_foreign_runtime_ids {
-  cuda = 1,
-  cuda_driver = 2,
-  opencl = 3,
-  sycl = 4,
-  hip = 5,
-  level_zero = 6,
-} omp_foreign_runtime_ids_t;
+/* Vendor defined values from OpenMP Additional Definitions document v2.1*/
+typedef enum omp_vendor_id {
+  omp_vendor_unknown = 0,
+  omp_vendor_amd = 1,
+  omp_vendor_arm = 2,
+  omp_vendor_bsc = 3,
+  omp_vendor_fujitsu = 4,
+  omp_vendor_gnu = 5,
+  omp_vendor_hpe = 6,
+  omp_vendor_ibm = 7,
+  omp_vendor_intel = 8,
+  omp_vendor_llvm = 9,
+  omp_vendor_nec = 10,
+  omp_vendor_nvidia = 11,
+  omp_vendor_ti = 12,
+  omp_vendor_last = 13
+} omp_vendor_id_t;
 
 ///} InteropAPI
 

offload/include/PerThreadTable.h

@@ -0,0 +1,114 @@
+//===-- PerThreadTable.h -- PerThread Storage Structure ----*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Table indexed with one entry per thread.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef OFFLOAD_PERTHREADTABLE_H
+#define OFFLOAD_PERTHREADTABLE_H
+
+#include <list>
+#include <memory>
+#include <mutex>
+
+// Using an STL container (such as std::vector) indexed by thread ID has
+// too many race conditions issues so we store each thread entry into a
+// thread_local variable.
+// T is the container type used to store the objects, e.g., std::vector,
+// std::set, etc. by each thread. O is the type of the stored objects e.g.,
+// omp_interop_val_t *, ...
+
+template <typename ContainerType, typename ObjectType> struct PerThreadTable {
+  using iterator = typename ContainerType::iterator;
+
+  struct PerThreadData {
+    size_t NElements = 0;
+    std::unique_ptr<ContainerType> ThEntry;
+  };
+
+  std::mutex Mtx;
+  std::list<std::shared_ptr<PerThreadData>> ThreadDataList;
+
+  // define default constructors, disable copy and move constructors
+  PerThreadTable() = default;
+  PerThreadTable(const PerThreadTable &) = delete;
+  PerThreadTable(PerThreadTable &&) = delete;
+  PerThreadTable &operator=(const PerThreadTable &) = delete;
+  PerThreadTable &operator=(PerThreadTable &&) = delete;
+  ~PerThreadTable() {
+    std::lock_guard<std::mutex> Lock(Mtx);
+    ThreadDataList.clear();
+  }
+
+private:
+  PerThreadData &getThreadData() {
+    static thread_local std::shared_ptr<PerThreadData> ThData = nullptr;
+    if (!ThData) {
+      ThData = std::make_shared<PerThreadData>();
+      std::lock_guard<std::mutex> Lock(Mtx);
+      ThreadDataList.push_back(ThData);
+    }
+    return *ThData;
+  }
+
+protected:
+  ContainerType &getThreadEntry() {
+    auto &ThData = getThreadData();
+    if (ThData.ThEntry)
+      return *ThData.ThEntry;
+    ThData.ThEntry = std::make_unique<ContainerType>();
+    return *ThData.ThEntry;
+  }
+
+  size_t &getThreadNElements() {
+    auto &ThData = getThreadData();
+    return ThData.NElements;
+  }
+
+public:
+  void add(ObjectType obj) {
+    auto &Entry = getThreadEntry();
+    auto &NElements = getThreadNElements();
+    NElements++;
+    Entry.add(obj);
+  }
+
+  iterator erase(iterator it) {
+    auto &Entry = getThreadEntry();
+    auto &NElements = getThreadNElements();
+    NElements--;
+    return Entry.erase(it);
+  }
+
+  size_t size() { return getThreadNElements(); }
+
+  // Iterators to traverse objects owned by
+  // the current thread
+  iterator begin() {
+    auto &Entry = getThreadEntry();
+    return Entry.begin();
+  }
+  iterator end() {
+    auto &Entry = getThreadEntry();
+    return Entry.end();
+  }
+
+  template <class F> void clear(F f) {
+    std::lock_guard<std::mutex> Lock(Mtx);
+    for (auto ThData : ThreadDataList) {
+      if (!ThData->ThEntry || ThData->NElements == 0)
+        continue;
+      ThData->ThEntry->clear(f);
+      ThData->NElements = 0;
+    }
+    ThreadDataList.clear();
+  }
+};
+
+#endif

offload/include/PluginManager.h

@@ -35,6 +35,8 @@
 #include <mutex>
 #include <string>
 
+#include "OpenMP/InteropAPI.h"
+
 using GenericPluginTy = llvm::omp::target::plugin::GenericPluginTy;
 
 /// Struct for the data required to handle plugins
@@ -88,6 +90,9 @@ struct PluginManager {
   HostPtrToTableMapTy HostPtrToTableMap;
   std::mutex TblMapMtx; ///< For HostPtrToTableMap
 
+  /// Table of cached implicit interop objects
+  InteropTblTy InteropTbl;
+
   // Work around for plugins that call dlopen on shared libraries that call
   // tgt_register_lib during their initialisation. Stash the pointers in a
   // vector until the plugins are all initialised and then register them.
@@ -185,5 +190,6 @@ void initRuntime();
 void deinitRuntime();
 
 extern PluginManager *PM;
-
+extern std::atomic<bool> RTLAlive; // Indicates if the RTL has been initialized
+extern std::atomic<int> RTLOngoingSyncs; // Counts ongoing external syncs
 #endif // OMPTARGET_PLUGIN_MANAGER_H

offload/include/Shared/APITypes.h

@@ -36,6 +36,7 @@ struct __tgt_device_image {
 struct __tgt_device_info {
   void *Context = nullptr;
   void *Device = nullptr;
+  void *Platform = nullptr;
 };
 
 /// This struct is a record of all the host code that may be offloaded to a