[Offload] Move RPC server handling to a dedicated thread (llvm#112988)

Summary:
Handling the RPC server requires running through list of jobs that the
device has requested to be done. Currently this is handled by the thread
that does the waiting for the kernel to finish. However, this is not
sound on NVIDIA architectures and only works for async launches in the
OpenMP model that uses helper threads.

However, we also don't want to have this thread doing work
unnnecessarily. For this reason we track the execution of kernels and
cause the thread to sleep via a condition variable (usually backed by
some kind of futex or other intelligent sleeping mechanism) so that the
thread will be idle while no kernels are running.

Author: jhuber6

offload/plugins-nextgen/amdgpu/src/rtl.cpp

@@ -1382,9 +1382,9 @@ struct AMDGPUSignalTy {
   }
 
   /// Wait until the signal gets a zero value.
-  Error wait(const uint64_t ActiveTimeout = 0, RPCServerTy *RPCServer = nullptr,
+  Error wait(const uint64_t ActiveTimeout = 0,
              GenericDeviceTy *Device = nullptr) const {
-    if (ActiveTimeout && !RPCServer) {
+    if (ActiveTimeout) {
       hsa_signal_value_t Got = 1;
       Got = hsa_signal_wait_scacquire(HSASignal, HSA_SIGNAL_CONDITION_EQ, 0,
                                       ActiveTimeout, HSA_WAIT_STATE_ACTIVE);
@@ -1393,14 +1393,11 @@ struct AMDGPUSignalTy {
     }
 
     // If there is an RPC device attached to this stream we run it as a server.
-    uint64_t Timeout = RPCServer ? 8192 : UINT64_MAX;
-    auto WaitState = RPCServer ? HSA_WAIT_STATE_ACTIVE : HSA_WAIT_STATE_BLOCKED;
+    uint64_t Timeout = UINT64_MAX;
+    auto WaitState = HSA_WAIT_STATE_BLOCKED;
     while (hsa_signal_wait_scacquire(HSASignal, HSA_SIGNAL_CONDITION_EQ, 0,
-                                     Timeout, WaitState) != 0) {
-      if (RPCServer && Device)
-        if (auto Err = RPCServer->runServer(*Device))
-          return Err;
-    }
+                                     Timeout, WaitState) != 0)
+      ;
     return Plugin::success();
   }
 
@@ -1895,11 +1892,6 @@ struct AMDGPUStreamTy {
   /// operation that was already finalized in a previous stream sycnhronize.
   uint32_t SyncCycle;
 
-  /// A pointer associated with an RPC server running on the given device. If
-  /// RPC is not being used this will be a null pointer. Otherwise, this
-  /// indicates that an RPC server is expected to be run on this stream.
-  RPCServerTy *RPCServer;
-
   /// Mutex to protect stream's management.
   mutable std::mutex Mutex;
 
@@ -2136,9 +2128,6 @@ struct AMDGPUStreamTy {
 
   hsa_queue_t *getHsaQueue() { return Queue->getHsaQueue(); }
 
-  /// Attach an RPC server to this stream.
-  void setRPCServer(RPCServerTy *Server) { RPCServer = Server; }
-
   /// Push a asynchronous kernel to the stream. The kernel arguments must be
   /// placed in a special allocation for kernel args and must keep alive until
   /// the kernel finalizes. Once the kernel is finished, the stream will release
@@ -2194,9 +2183,30 @@ struct AMDGPUStreamTy {
 
     // Push the kernel with the output signal and an input signal (optional)
     DP("Using Queue: %p with HSA Queue: %p\n", Queue, Queue->getHsaQueue());
-    return Queue->pushKernelLaunch(Kernel, KernelArgs, NumThreads, NumBlocks,
-                                   GroupSize, StackSize, OutputSignal,
-                                   InputSignal);
+    // If we are running an RPC server we want to wake up the server thread
+    // whenever there is a kernel running and let it sleep otherwise.
+    if (Device.getRPCServer())
+      Device.Plugin.getRPCServer().Thread->notify();
+
+    // Push the kernel with the output signal and an input signal (optional)
+    if (auto Err = Queue->pushKernelLaunch(Kernel, KernelArgs, NumThreads,
+                                           NumBlocks, GroupSize, StackSize,
+                                           OutputSignal, InputSignal))
+      return Err;
+
+    // Register a callback to indicate when the kernel is complete.
+    if (Device.getRPCServer()) {
+      if (auto Err = Slots[Curr].schedCallback(
+              [](void *Data) -> llvm::Error {
+                GenericPluginTy &Plugin =
+                    *reinterpret_cast<GenericPluginTy *>(Data);
+                Plugin.getRPCServer().Thread->finish();
+                return Error::success();
+              },
+              &Device.Plugin))
+        return Err;
+    }
+    return Plugin::success();
   }
 
   /// Push an asynchronous memory copy between pinned memory buffers.
@@ -2268,9 +2278,8 @@ struct AMDGPUStreamTy {
 
     // Wait for kernel to finish before scheduling the asynchronous copy.
     if (UseSyncCopyBack && InputSignal && InputSignal->load())
-      if (auto Err = InputSignal->wait(StreamBusyWaitMicroseconds, RPCServer, &Device))
+      if (auto Err = InputSignal->wait(StreamBusyWaitMicroseconds, &Device))
         return Err;
-
 #ifdef OMPT_SUPPORT
 
     if (OmptInfo) {
@@ -2457,8 +2466,8 @@ struct AMDGPUStreamTy {
       return Plugin::success();
 
     // Wait until all previous operations on the stream have completed.
-    if (auto Err = Slots[last()].Signal->wait(StreamBusyWaitMicroseconds,
-                                              RPCServer, &Device))
+    if (auto Err =
+            Slots[last()].Signal->wait(StreamBusyWaitMicroseconds, &Device))
       return Err;
 
     // Reset the stream and perform all pending post actions.
@@ -4701,7 +4710,7 @@ AMDGPUStreamTy::AMDGPUStreamTy(AMDGPUDeviceTy &Device)
     : Agent(Device.getAgent()), Queue(nullptr),
       SignalManager(Device.getSignalManager()), Device(Device),
       // Initialize the std::deque with some empty positions.
-      Slots(32), NextSlot(0), SyncCycle(0), RPCServer(nullptr),
+      Slots(32), NextSlot(0), SyncCycle(0),
       StreamBusyWaitMicroseconds(Device.getStreamBusyWaitMicroseconds()),
       UseMultipleSdmaEngines(Device.useMultipleSdmaEngines()),
       UseSyncCopyBack(Device.syncCopyBack()) {}
@@ -5117,10 +5126,6 @@ Error AMDGPUKernelTy::launchImpl(GenericDeviceTy &GenericDevice,
     DP("No hostrpc buffer or service thread required\n");
   }
 
-  // If this kernel requires an RPC server we attach its pointer to the stream.
-  if (GenericDevice.getRPCServer())
-    Stream->setRPCServer(GenericDevice.getRPCServer());
-
   // Only COV5 implicitargs needs to be set. COV4 implicitargs are not used.
   if (ImplArgs &&
       getImplicitArgsSize() == sizeof(hsa_utils::AMDGPUImplicitArgsTy)) {

offload/plugins-nextgen/common/include/PluginInterface.h

@@ -797,12 +797,6 @@ struct GenericDeviceTy : public DeviceAllocatorTy {
   /// Setup the global device memory pool, if the plugin requires one.
   Error setupDeviceMemoryPool(GenericPluginTy &Plugin, DeviceImageTy &Image,
                               uint64_t PoolSize);
-
-  // Setup the RPC server for this device if needed. This may not run on some
-  // plugins like the CPU targets. By default, it will not be executed so it is
-  // up to the target to override this using the shouldSetupRPCServer function.
-  Error setupRPCServer(GenericPluginTy &Plugin, DeviceImageTy &Image);
-
   /// Synchronize the current thread with the pending operations on the
   /// __tgt_async_info structure.
   Error synchronize(__tgt_async_info *AsyncInfo);

offload/plugins-nextgen/common/include/RPC.h

@@ -19,7 +19,11 @@
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/Support/Error.h"
 
+#include <atomic>
+#include <condition_variable>
 #include <cstdint>
+#include <mutex>
+#include <thread>
 
 namespace llvm::omp::target {
 namespace plugin {
@@ -37,6 +41,12 @@ struct RPCServerTy {
   /// Initializes the handles to the number of devices we may need to service.
   RPCServerTy(plugin::GenericPluginTy &Plugin);
 
+  /// Deinitialize the associated memory and resources.
+  llvm::Error shutDown();
+
+  /// Initialize the worker thread.
+  llvm::Error startThread();
+
   /// Check if this device image is using an RPC server. This checks for the
   /// precense of an externally visible symbol in the device image that will
   /// be present whenever RPC code is called.
@@ -51,17 +61,77 @@ struct RPCServerTy {
                          plugin::GenericGlobalHandlerTy &Handler,
                          plugin::DeviceImageTy &Image);
 
-  /// Runs the RPC server associated with the \p Device until the pending work
-  /// is cleared.
-  llvm::Error runServer(plugin::GenericDeviceTy &Device);
-
   /// Deinitialize the RPC server for the given device. This will free the
   /// memory associated with the k
   llvm::Error deinitDevice(plugin::GenericDeviceTy &Device);
 
 private:
   /// Array from this device's identifier to its attached devices.
-  llvm::SmallVector<void *> Buffers;
+  std::unique_ptr<void *[]> Buffers;
+
+  /// Array of associated devices. These must be alive as long as the server is.
+  std::unique_ptr<plugin::GenericDeviceTy *[]> Devices;
+
+  /// A helper class for running the user thread that handles the RPC interface.
+  /// Because we only need to check the RPC server while any kernels are
+  /// working, we track submission / completion events to allow the thread to
+  /// sleep when it is not needed.
+  struct ServerThread {
+    std::thread Worker;
+
+    /// A boolean indicating whether or not the worker thread should continue.
+    std::atomic<bool> Running;
+
+    /// The number of currently executing kernels across all devices that need
+    /// the server thread to be running.
+    std::atomic<uint32_t> NumUsers;
+
+    /// The condition variable used to suspend the thread if no work is needed.
+    std::condition_variable CV;
+    std::mutex Mutex;
+
+    /// A reference to all the RPC interfaces that the server is handling.
+    llvm::ArrayRef<void *> Buffers;
+
+    /// A reference to the associated generic device for the buffer.
+    llvm::ArrayRef<plugin::GenericDeviceTy *> Devices;
+
+    /// Initialize the worker thread to run in the background.
+    ServerThread(void *Buffers[], plugin::GenericDeviceTy *Devices[],
+                 size_t Length)
+        : Running(true), NumUsers(0), CV(), Mutex(), Buffers(Buffers, Length),
+          Devices(Devices, Length) {}
+
+    ~ServerThread() { assert(!Running && "Thread not shut down explicitly\n"); }
+
+    /// Notify the worker thread that there is a user that needs it.
+    void notify() {
+      std::lock_guard<decltype(Mutex)> Lock(Mutex);
+      NumUsers.fetch_add(1, std::memory_order_relaxed);
+      CV.notify_all();
+    }
+
+    /// Indicate that one of the dependent users has finished.
+    void finish() {
+      [[maybe_unused]] uint32_t Old =
+          NumUsers.fetch_sub(1, std::memory_order_relaxed);
+      assert(Old > 0 && "Attempt to signal finish with no pending work");
+    }
+
+    /// Destroy the worker thread and wait.
+    void shutDown();
+
+    /// Initialize the worker thread.
+    void startThread();
+
+    /// Run the server thread to continuously check the RPC interface for work
+    /// to be done for the device.
+    void run();
+  };
+
+public:
+  /// Pointer to the server thread instance.
+  std::unique_ptr<ServerThread> Thread;
 };
 
 } // namespace llvm::omp::target

offload/plugins-nextgen/common/src/PluginInterface.cpp

@@ -1132,10 +1132,6 @@ GenericDeviceTy::loadBinary(GenericPluginTy &Plugin,
     } else if (auto Err = setupDeviceMemoryPool(Plugin, *Image, HeapSize))
       return std::move(Err);
   }
-
-  if (auto Err = setupRPCServer(Plugin, *Image))
-    return std::move(Err);
-
 #ifdef OMPT_SUPPORT
   if (ompt::Initialized) {
     size_t Bytes =
@@ -1249,30 +1245,6 @@ Error GenericDeviceTy::setupDeviceMemoryPool(GenericPluginTy &Plugin,
   return GHandler.writeGlobalToDevice(*this, Image, DevEnvGlobal);
 }
 
-Error GenericDeviceTy::setupRPCServer(GenericPluginTy &Plugin,
-                                      DeviceImageTy &Image) {
-  // The plugin either does not need an RPC server or it is unavailible.
-  if (!shouldSetupRPCServer())
-    return Plugin::success();
-
-  // Check if this device needs to run an RPC server.
-  RPCServerTy &Server = Plugin.getRPCServer();
-  auto UsingOrErr =
-      Server.isDeviceUsingRPC(*this, Plugin.getGlobalHandler(), Image);
-  if (!UsingOrErr)
-    return UsingOrErr.takeError();
-
-  if (!UsingOrErr.get())
-    return Plugin::success();
-
-  if (auto Err = Server.initDevice(*this, Plugin.getGlobalHandler(), Image))
-    return Err;
-
-  RPCServer = &Server;
-  DP("Running an RPC server on device %d\n", getDeviceId());
-  return Plugin::success();
-}
-
 Error PinnedAllocationMapTy::insertEntry(void *HstPtr, void *DevAccessiblePtr,
                                          size_t Size, bool ExternallyLocked) {
   // Insert the new entry into the map.
@@ -1892,8 +1864,11 @@ Error GenericPluginTy::deinit() {
     delete GlobalHandler;
 
 #if RPC_FIXME
-  if (RPCServer)
+  if (RPCServer) {
+    if (Error Err = RPCServer->shutDown())
+      return Err;
     delete RPCServer;
+  }
 #endif
 
   if (RecordReplay)