[SYCL][CUDA][PI] Improve performance of event synchronization (#6224)

Improve performance of event synchronization by reducing the number of calls to cuStreamWaitEvent. This call is now skipped for the stream, the event is coming from. Also when enqueueing a new command with a dependency on a previous one an attempt to use the same stream will be made, so both can be waited on by only one call to cuStreamWaitEvent.

Author: t4c1

sycl/plugins/cuda/pi_cuda.cpp

@@ -306,7 +306,11 @@ pi_result enqueueEventsWait(pi_queue command_queue, CUstream stream,
     auto result = forLatestEvents(
         event_wait_list, num_events_in_wait_list,
         [stream](pi_event event) -> pi_result {
-          return PI_CHECK_ERROR(cuStreamWaitEvent(stream, event->get(), 0));
+          if (event->get_stream() == stream) {
+            return PI_SUCCESS;
+          } else {
+            return PI_CHECK_ERROR(cuStreamWaitEvent(stream, event->get(), 0));
+          }
         });
 
     if (result != PI_SUCCESS) {
@@ -373,18 +377,58 @@ pi_result cuda_piEventRetain(pi_event event);
 
 /// \endcond
 
-CUstream _pi_queue::get_next_compute_stream() {
-  if (num_compute_streams_ < compute_streams_.size()) {
-    // the check above is for performance - so as not to lock mutex every time
-    std::lock_guard<std::mutex> guard(compute_stream_mutex_);
-    // The second check is done after mutex is locked so other threads can not
-    // change num_compute_streams_ after that
+CUstream _pi_queue::get_next_compute_stream(pi_uint32 *stream_token) {
+  pi_uint32 stream_i;
+  while (true) {
     if (num_compute_streams_ < compute_streams_.size()) {
-      PI_CHECK_ERROR(
-          cuStreamCreate(&compute_streams_[num_compute_streams_++], flags_));
+      // the check above is for performance - so as not to lock mutex every time
+      std::lock_guard<std::mutex> guard(compute_stream_mutex_);
+      // The second check is done after mutex is locked so other threads can not
+      // change num_compute_streams_ after that
+      if (num_compute_streams_ < compute_streams_.size()) {
+        PI_CHECK_ERROR(
+            cuStreamCreate(&compute_streams_[num_compute_streams_++], flags_));
+      }
+    }
+    stream_i = compute_stream_idx_++;
+    // if a stream has been reused before it was next selected round-robin
+    // fashion, we want to delay its next use and instead select another one
+    // that is more likely to have completed all the enqueued work.
+    if (delay_compute_[stream_i % compute_streams_.size()]) {
+      delay_compute_[stream_i % compute_streams_.size()] = false;
+    } else {
+      break;
     }
   }
-  return compute_streams_[compute_stream_idx_++ % compute_streams_.size()];
+  if (stream_token) {
+    *stream_token = stream_i;
+  }
+  return compute_streams_[stream_i % compute_streams_.size()];
+}
+
+CUstream _pi_queue::get_next_compute_stream(pi_uint32 num_events_in_wait_list,
+                                            const pi_event *event_wait_list,
+                                            _pi_stream_guard &guard,
+                                            pi_uint32 *stream_token) {
+  for (pi_uint32 i = 0; i < num_events_in_wait_list; i++) {
+    pi_uint32 token = event_wait_list[i]->get_stream_token();
+    if (event_wait_list[i]->get_queue() == this && can_reuse_stream(token)) {
+      std::unique_lock<std::mutex> compute_sync_guard(
+          compute_stream_sync_mutex_);
+      // redo the check after lock to avoid data races on
+      // last_sync_compute_streams_
+      if (can_reuse_stream(token)) {
+        delay_compute_[token % delay_compute_.size()] = true;
+        if (stream_token) {
+          *stream_token = token;
+        }
+        guard = _pi_stream_guard{std::move(compute_sync_guard)};
+        return event_wait_list[i]->get_stream();
+      }
+    }
+  }
+  guard = {};
+  return get_next_compute_stream(stream_token);
 }
 
 CUstream _pi_queue::get_next_transfer_stream() {
@@ -405,9 +449,10 @@ CUstream _pi_queue::get_next_transfer_stream() {
 }
 
 _pi_event::_pi_event(pi_command_type type, pi_context context, pi_queue queue,
-                     CUstream stream)
+                     CUstream stream, pi_uint32 stream_token)
     : commandType_{type}, refCount_{1}, hasBeenWaitedOn_{false},
-      isRecorded_{false}, isStarted_{false}, evEnd_{nullptr}, evStart_{nullptr},
+      isRecorded_{false}, isStarted_{false},
+      streamToken_{stream_token}, evEnd_{nullptr}, evStart_{nullptr},
       evQueued_{nullptr}, queue_{queue}, stream_{stream}, context_{context} {
 
   bool profilingEnabled = queue_->properties_ & PI_QUEUE_PROFILING_ENABLE;
@@ -2948,7 +2993,10 @@ pi_result cuda_piEnqueueKernelLaunch(
 
     std::unique_ptr<_pi_event> retImplEv{nullptr};
 
-    CUstream cuStream = command_queue->get_next_compute_stream();
+    pi_uint32 stream_token;
+    _pi_stream_guard guard;
+    CUstream cuStream = command_queue->get_next_compute_stream(
+        num_events_in_wait_list, event_wait_list, guard, &stream_token);
     CUfunction cuFunc = kernel->get();
 
     retError = enqueueEventsWait(command_queue, cuStream,
@@ -2973,8 +3021,9 @@ pi_result cuda_piEnqueueKernelLaunch(
     auto &argIndices = kernel->get_arg_indices();
 
     if (event) {
-      retImplEv = std::unique_ptr<_pi_event>(_pi_event::make_native(
-          PI_COMMAND_TYPE_NDRANGE_KERNEL, command_queue, cuStream));
+      retImplEv = std::unique_ptr<_pi_event>(
+          _pi_event::make_native(PI_COMMAND_TYPE_NDRANGE_KERNEL, command_queue,
+                                 cuStream, stream_token));
       retImplEv->start();
     }
 
@@ -3809,7 +3858,12 @@ pi_result cuda_piEnqueueEventsWaitWithBarrier(pi_queue command_queue,
       auto result =
           forLatestEvents(event_wait_list, num_events_in_wait_list,
                           [command_queue](pi_event event) -> pi_result {
-                            return enqueueEventWait(command_queue, event);
+                            if (event->get_queue()->has_been_synchronized(
+                                    event->get_stream_token())) {
+                              return PI_SUCCESS;
+                            } else {
+                              return enqueueEventWait(command_queue, event);
+                            }
                           });
 
       if (result != PI_SUCCESS) {
@@ -3818,8 +3872,12 @@ pi_result cuda_piEnqueueEventsWaitWithBarrier(pi_queue command_queue,
     }
 
     if (event) {
+      pi_uint32 stream_token;
+      _pi_stream_guard guard;
+      CUstream cuStream = command_queue->get_next_compute_stream(
+          num_events_in_wait_list, event_wait_list, guard, &stream_token);
       *event = _pi_event::make_native(PI_COMMAND_TYPE_MARKER, command_queue,
-                                      command_queue->get_next_compute_stream());
+                                      cuStream, stream_token);
       (*event)->start();
       (*event)->record();
     }
@@ -4879,12 +4937,15 @@ pi_result cuda_piextUSMEnqueueMemset(pi_queue queue, void *ptr, pi_int32 value,
 
   try {
     ScopedContext active(queue->get_context());
-    CUstream cuStream = queue->get_next_compute_stream();
+    pi_uint32 stream_token;
+    _pi_stream_guard guard;
+    CUstream cuStream = queue->get_next_compute_stream(
+        num_events_in_waitlist, events_waitlist, guard, &stream_token);
     result = enqueueEventsWait(queue, cuStream, num_events_in_waitlist,
                                events_waitlist);
     if (event) {
       event_ptr = std::unique_ptr<_pi_event>(_pi_event::make_native(
-          PI_COMMAND_TYPE_MEM_BUFFER_FILL, queue, cuStream));
+          PI_COMMAND_TYPE_MEM_BUFFER_FILL, queue, cuStream, stream_token));
       event_ptr->start();
     }
     result = PI_CHECK_ERROR(cuMemsetD8Async(

sycl/plugins/cuda/pi_cuda.hpp

@@ -55,6 +55,8 @@ pi_result cuda_piKernelGetGroupInfo(pi_kernel kernel, pi_device device,
 /// \endcond
 }
 
+using _pi_stream_guard = std::unique_lock<std::mutex>;
+
 /// A PI platform stores all known PI devices,
 ///  in the CUDA plugin this is just a vector of
 ///  available devices since initialization is done
@@ -387,6 +389,11 @@ struct _pi_queue {
 
   std::vector<native_type> compute_streams_;
   std::vector<native_type> transfer_streams_;
+  // delay_compute_ keeps track of which streams have been recently reused and
+  // their next use should be delayed. If a stream has been recently reused it
+  // will be skipped the next time it would be selected round-robin style. When
+  // skipped, its delay flag is cleared.
+  std::vector<bool> delay_compute_;
   _pi_context *context_;
   _pi_device *device_;
   pi_queue_properties properties_;
@@ -399,6 +406,10 @@ struct _pi_queue {
   unsigned int last_sync_compute_streams_;
   unsigned int last_sync_transfer_streams_;
   unsigned int flags_;
+  // When compute_stream_sync_mutex_ and compute_stream_mutex_ both need to be
+  // locked at the same time, compute_stream_sync_mutex_ should be locked first
+  // to avoid deadlocks
+  std::mutex compute_stream_sync_mutex_;
   std::mutex compute_stream_mutex_;
   std::mutex transfer_stream_mutex_;
   bool has_ownership_;
@@ -408,7 +419,8 @@ struct _pi_queue {
             _pi_device *device, pi_queue_properties properties,
             unsigned int flags, bool backend_owns = true)
       : compute_streams_{std::move(compute_streams)},
-        transfer_streams_{std::move(transfer_streams)}, context_{context},
+        transfer_streams_{std::move(transfer_streams)},
+        delay_compute_(compute_streams_.size(), false), context_{context},
         device_{device}, properties_{properties}, refCount_{1}, eventCount_{0},
         compute_stream_idx_{0}, transfer_stream_idx_{0},
         num_compute_streams_{0}, num_transfer_streams_{0},
@@ -425,10 +437,47 @@ struct _pi_queue {
 
   // get_next_compute/transfer_stream() functions return streams from
   // appropriate pools in round-robin fashion
-  native_type get_next_compute_stream();
+  native_type get_next_compute_stream(pi_uint32 *stream_token = nullptr);
+  // this overload tries select a stream that was used by one of dependancies.
+  // If that is not possible returns a new stream. If a stream is reused it
+  // returns a lock that needs to remain locked as long as the stream is in use
+  native_type get_next_compute_stream(pi_uint32 num_events_in_wait_list,
+                                      const pi_event *event_wait_list,
+                                      _pi_stream_guard &guard,
+                                      pi_uint32 *stream_token = nullptr);
   native_type get_next_transfer_stream();
   native_type get() { return get_next_compute_stream(); };
 
+  bool has_been_synchronized(pi_uint32 stream_token) {
+    // stream token not associated with one of the compute streams
+    if (stream_token == std::numeric_limits<pi_uint32>::max()) {
+      return false;
+    }
+    return last_sync_compute_streams_ >= stream_token;
+  }
+
+  bool can_reuse_stream(pi_uint32 stream_token) {
+    // stream token not associated with one of the compute streams
+    if (stream_token == std::numeric_limits<pi_uint32>::max()) {
+      return true;
+    }
+    // If the command represented by the stream token was not the last command
+    // enqueued to the stream we can not reuse the stream - we need to allow for
+    // commands enqueued after it and the one we are about to enqueue to run
+    // concurrently
+    bool is_last_command =
+        (compute_stream_idx_ - stream_token) <= compute_streams_.size();
+    // If there was a barrier enqueued to the queue after the command
+    // represented by the stream token we should not reuse the stream, as we can
+    // not take that stream into account for the bookkeeping for the next
+    // barrier - such a stream would not be synchronized with. Performance-wise
+    // it does not matter that we do not reuse the stream, as the work
+    // represented by the stream token is guaranteed to be complete by the
+    // barrier before any work we are about to enqueue to the stream will start,
+    // so the event does not need to be synchronized with.
+    return is_last_command && !has_been_synchronized(stream_token);
+  }
+
   template <typename T> void for_each_stream(T &&f) {
     {
       std::lock_guard<std::mutex> compute_guard(compute_stream_mutex_);
@@ -451,30 +500,39 @@ struct _pi_queue {
   }
 
   template <typename T> void sync_streams(T &&f) {
-    auto sync = [&f](const std::vector<CUstream> &streams, unsigned int start,
-                     unsigned int stop) {
+    auto sync_compute = [&f, &streams = compute_streams_,
+                         &delay = delay_compute_](unsigned int start,
+                                                  unsigned int stop) {
+      for (unsigned int i = start; i < stop; i++) {
+        f(streams[i]);
+        delay[i] = false;
+      }
+    };
+    auto sync_transfer = [&f, &streams = transfer_streams_](unsigned int start,
+                                                            unsigned int stop) {
       for (unsigned int i = start; i < stop; i++) {
         f(streams[i]);
       }
     };
     {
       unsigned int size = static_cast<unsigned int>(compute_streams_.size());
+      std::lock_guard compute_sync_guard(compute_stream_sync_mutex_);
       std::lock_guard<std::mutex> compute_guard(compute_stream_mutex_);
       unsigned int start = last_sync_compute_streams_;
       unsigned int end = num_compute_streams_ < size
                              ? num_compute_streams_
                              : compute_stream_idx_.load();
       last_sync_compute_streams_ = end;
       if (end - start >= size) {
-        sync(compute_streams_, 0, size);
+        sync_compute(0, size);
       } else {
         start %= size;
         end %= size;
         if (start < end) {
-          sync(compute_streams_, start, end);
+          sync_compute(start, end);
         } else {
-          sync(compute_streams_, start, size);
-          sync(compute_streams_, 0, end);
+          sync_compute(start, size);
+          sync_compute(0, end);
         }
       }
     }
@@ -488,15 +546,15 @@ struct _pi_queue {
                                : transfer_stream_idx_.load();
         last_sync_transfer_streams_ = end;
         if (end - start >= size) {
-          sync(transfer_streams_, 0, size);
+          sync_transfer(0, size);
         } else {
           start %= size;
           end %= size;
           if (start < end) {
-            sync(transfer_streams_, start, end);
+            sync_transfer(start, end);
           } else {
-            sync(transfer_streams_, start, size);
-            sync(transfer_streams_, 0, end);
+            sync_transfer(start, size);
+            sync_transfer(0, end);
           }
         }
       }
@@ -538,6 +596,8 @@ struct _pi_event {
 
   CUstream get_stream() const noexcept { return stream_; }
 
+  pi_uint32 get_stream_token() const noexcept { return streamToken_; }
+
   pi_command_type get_command_type() const noexcept { return commandType_; }
 
   pi_uint32 get_reference_count() const noexcept { return refCount_; }
@@ -581,9 +641,11 @@ struct _pi_event {
   pi_uint64 get_end_time() const;
 
   // construct a native CUDA. This maps closely to the underlying CUDA event.
-  static pi_event make_native(pi_command_type type, pi_queue queue,
-                              CUstream stream) {
-    return new _pi_event(type, queue->get_context(), queue, stream);
+  static pi_event
+  make_native(pi_command_type type, pi_queue queue, CUstream stream,
+              pi_uint32 stream_token = std::numeric_limits<pi_uint32>::max()) {
+    return new _pi_event(type, queue->get_context(), queue, stream,
+                         stream_token);
   }
 
   pi_result release();
@@ -594,7 +656,7 @@ struct _pi_event {
   // This constructor is private to force programmers to use the make_native /
   // make_user static members in order to create a pi_event for CUDA.
   _pi_event(pi_command_type type, pi_context context, pi_queue queue,
-            CUstream stream);
+            CUstream stream, pi_uint32 stream_token);
 
   pi_command_type commandType_; // The type of command associated with event.
 
@@ -610,6 +672,7 @@ struct _pi_event {
                     // PI event has started or not
                     //
 
+  pi_uint32 streamToken_;
   pi_uint32 eventId_; // Queue identifier of the event.
 
   native_type evEnd_; // CUDA event handle. If this _pi_event represents a user