[Offload] Move RPC server handling to a dedicated thread

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

@@ -626,9 +626,9 @@ 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);
@@ -637,14 +637,11 @@ 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();
   }
 
@@ -1052,11 +1049,6 @@ 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;
 
@@ -1236,9 +1228,6 @@ struct AMDGPUStreamTy {
   /// Deinitialize the stream's signals.
   Error deinit() { return Plugin::success(); }
 
-  /// 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
@@ -1266,10 +1255,30 @@ struct AMDGPUStreamTy {
     if (auto Err = Slots[Curr].schedReleaseBuffer(KernelArgs, MemoryManager))
       return Err;
 
+    // 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)
-    return Queue->pushKernelLaunch(Kernel, KernelArgs, NumThreads, NumBlocks,
-                                   GroupSize, StackSize, OutputSignal,
-                                   InputSignal);
+    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.
@@ -1479,8 +1488,8 @@ 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.
@@ -3024,7 +3033,7 @@ 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()) {}
 
@@ -3377,10 +3386,6 @@ Error AMDGPUKernelTy::launchImpl(GenericDeviceTy &GenericDevice,
   if (auto Err = AMDGPUDevice.getStream(AsyncInfoWrapper, Stream))
     return Err;
 
-  // 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/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 +41,9 @@ 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();
+
   /// 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 +58,61 @@ 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<uintptr_t> Handles;
+  std::unique_ptr<std::atomic<uintptr_t>[]> Handles;
+
+  /// A helper class for running the user thread that handles
+  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<std::atomic<uintptr_t>> Handles;
+
+    /// Initialize the worker thread to run in the background.
+    ServerThread(std::atomic<uintptr_t> Handles[], size_t 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();
+
+    /// 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

@@ -1624,8 +1624,11 @@ Error GenericPluginTy::deinit() {
   if (GlobalHandler)
     delete GlobalHandler;
 
-  if (RPCServer)
+  if (RPCServer) {
+    if (Error Err = RPCServer->shutDown())
+      return Err;
     delete RPCServer;
+  }
 
   if (RecordReplay)
     delete RecordReplay;

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

@@ -21,8 +21,64 @@ using namespace llvm;
 using namespace omp;
 using namespace target;
 
+void RPCServerTy::ServerThread::shutDown() {
+#ifdef LIBOMPTARGET_RPC_SUPPORT
+  {
+    std::lock_guard<decltype(Mutex)> Lock(Mutex);
+    Running.store(false, std::memory_order_release);
+    CV.notify_all();
+  }
+  if (Worker.joinable())
+    Worker.join();
+#endif
+}
+
+void RPCServerTy::ServerThread::run() {
+#ifdef LIBOMPTARGET_RPC_SUPPORT
+  while (Running.load(std::memory_order_acquire)) {
+    std::unique_lock<decltype(Mutex)> Lock(Mutex);
+    CV.wait(Lock, [&]() {
+      return NumUsers.load(std::memory_order_acquire) > 0 ||
+             !Running.load(std::memory_order_acquire);
+    });
+
+    Lock.unlock();
+
+    if (!Running.load(std::memory_order_seq_cst))
+      return;
+
+    while (NumUsers.load(std::memory_order_relaxed) > 0 &&
+           Running.load(std::memory_order_relaxed)) {
+      for (const auto &Handle : Handles) {
+        rpc_device_t RPCDevice{Handle};
+        [[maybe_unused]] rpc_status_t Err = rpc_handle_server(RPCDevice);
+        assert(Err == RPC_STATUS_SUCCESS &&
+               "Checking the RPC server should not fail");
+      }
+    }
+  }
+#endif
+}
+
+RPCServerTy::ServerThread::ServerThread(std::atomic<uintptr_t> Handles[],
+                                        size_t Length)
+    : Running(true), NumUsers(0), CV(), Mutex(), Handles(Handles, Length) {
+#ifdef LIBOMPTARGET_RPC_SUPPORT
+  Worker = std::thread([this]() { run(); });
+#endif
+}
+
 RPCServerTy::RPCServerTy(plugin::GenericPluginTy &Plugin)
-    : Handles(Plugin.getNumDevices()) {}
+    : Handles(
+          std::make_unique<std::atomic<uintptr_t>[]>(Plugin.getNumDevices())),
+      Thread(new ServerThread(Handles.get(), Plugin.getNumDevices())) {}
+
+llvm::Error RPCServerTy::shutDown() {
+#ifdef LIBOMPTARGET_RPC_SUPPORT
+  Thread->shutDown();
+  return Error::success();
+#endif
+}
 
 llvm::Expected<bool>
 RPCServerTy::isDeviceUsingRPC(plugin::GenericDeviceTy &Device,
@@ -109,17 +165,6 @@ Error RPCServerTy::initDevice(plugin::GenericDeviceTy &Device,
   return Error::success();
 }
 
-Error RPCServerTy::runServer(plugin::GenericDeviceTy &Device) {
-#ifdef LIBOMPTARGET_RPC_SUPPORT
-  rpc_device_t RPCDevice{Handles[Device.getDeviceId()]};
-  if (rpc_status_t Err = rpc_handle_server(RPCDevice))
-    return plugin::Plugin::error(
-        "Error while running RPC server on device %d: %d", Device.getDeviceId(),
-        Err);
-#endif
-  return Error::success();
-}
-
 Error RPCServerTy::deinitDevice(plugin::GenericDeviceTy &Device) {
 #ifdef LIBOMPTARGET_RPC_SUPPORT
   rpc_device_t RPCDevice{Handles[Device.getDeviceId()]};

offload/plugins-nextgen/cuda/dynamic_cuda/cuda.cpp

@@ -63,6 +63,7 @@ DLWRAP(cuStreamCreate, 2)
 DLWRAP(cuStreamDestroy, 1)
 DLWRAP(cuStreamSynchronize, 1)
 DLWRAP(cuStreamQuery, 1)
+DLWRAP(cuStreamAddCallback, 4)
 DLWRAP(cuCtxSetCurrent, 1)
 DLWRAP(cuDevicePrimaryCtxRelease, 1)
 DLWRAP(cuDevicePrimaryCtxGetState, 3)

offload/plugins-nextgen/cuda/dynamic_cuda/cuda.h

@@ -286,6 +286,8 @@ static inline void *CU_LAUNCH_PARAM_END = (void *)0x00;
 static inline void *CU_LAUNCH_PARAM_BUFFER_POINTER = (void *)0x01;
 static inline void *CU_LAUNCH_PARAM_BUFFER_SIZE = (void *)0x02;
 
+typedef void (*CUstreamCallback)(CUstream, CUresult, void *);
+
 CUresult cuCtxGetDevice(CUdevice *);
 CUresult cuDeviceGet(CUdevice *, int);
 CUresult cuDeviceGetAttribute(int *, CUdevice_attribute, CUdevice);
@@ -326,6 +328,7 @@ CUresult cuStreamCreate(CUstream *, unsigned);
 CUresult cuStreamDestroy(CUstream);
 CUresult cuStreamSynchronize(CUstream);
 CUresult cuStreamQuery(CUstream);
+CUresult cuStreamAddCallback(CUstream, CUstreamCallback, void *, unsigned int);
 CUresult cuCtxSetCurrent(CUcontext);
 CUresult cuDevicePrimaryCtxRelease(CUdevice);
 CUresult cuDevicePrimaryCtxGetState(CUdevice, unsigned *, int *);

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

@@ -632,15 +632,7 @@ struct CUDADeviceTy : public GenericDeviceTy {
     CUresult Res;
     // If we have an RPC server running on this device we will continuously
     // query it for work rather than blocking.
-    if (!getRPCServer()) {
-      Res = cuStreamSynchronize(Stream);
-    } else {
-      do {
-        Res = cuStreamQuery(Stream);
-        if (auto Err = getRPCServer()->runServer(*this))
-          return Err;
-      } while (Res == CUDA_ERROR_NOT_READY);
-    }
+    Res = cuStreamSynchronize(Stream);
 
     // Once the stream is synchronized, return it to stream pool and reset
     // AsyncInfo. This is to make sure the synchronization only works for its
@@ -825,17 +817,6 @@ struct CUDADeviceTy : public GenericDeviceTy {
     if (auto Err = getStream(AsyncInfoWrapper, Stream))
       return Err;
 
-    // If there is already pending work on the stream it could be waiting for
-    // someone to check the RPC server.
-    if (auto *RPCServer = getRPCServer()) {
-      CUresult Res = cuStreamQuery(Stream);
-      while (Res == CUDA_ERROR_NOT_READY) {
-        if (auto Err = RPCServer->runServer(*this))
-          return Err;
-        Res = cuStreamQuery(Stream);
-      }
-    }
-
     CUresult Res = cuMemcpyDtoHAsync(HstPtr, (CUdeviceptr)TgtPtr, Size, Stream);
     return Plugin::check(Res, "Error in cuMemcpyDtoHAsync: %s");
   }
@@ -1294,10 +1275,26 @@ Error CUDAKernelTy::launchImpl(GenericDeviceTy &GenericDevice,
                     reinterpret_cast<void *>(&LaunchParams.Size),
                     CU_LAUNCH_PARAM_END};
 
+  // 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 (GenericDevice.getRPCServer())
+    GenericDevice.Plugin.getRPCServer().Thread->notify();
+
   CUresult Res = cuLaunchKernel(Func, NumBlocks, /*gridDimY=*/1,
                                 /*gridDimZ=*/1, NumThreads,
                                 /*blockDimY=*/1, /*blockDimZ=*/1,
                                 MaxDynCGroupMem, Stream, nullptr, Config);
+
+  // Register a callback to indicate when the kernel is complete.
+  if (GenericDevice.getRPCServer())
+    cuStreamAddCallback(
+        Stream,
+        [](CUstream Stream, CUresult Status, void *Data) {
+          GenericPluginTy &Plugin = *reinterpret_cast<GenericPluginTy *>(Data);
+          Plugin.getRPCServer().Thread->finish();
+        },
+        &GenericDevice.Plugin, /*flags=*/0);
+
   return Plugin::check(Res, "Error in cuLaunchKernel for '%s': %s", getName());
 }