Fixes

Signed-off-by: Vladimir Paramuzov <[email protected]>

src/plugins/intel_gpu/include/intel_gpu/graph/network.hpp

@@ -36,7 +36,7 @@ struct network_output {
         // TODO: in_order queue doesn't create proper output event in some cases which leads to syncronization issues with user app
         // So call finish for associated stream to enusre that the output data is ready.
         if (do_sync) {
-            if (_stream->get_queue_type() == QueueTypes::in_order) {
+            if (_stream->get_queue_type() == QueueTypes::in_order || !_event) {
                 _stream->finish();
             } else {
                 _event->wait();

src/plugins/intel_gpu/src/graph/impls/cpu/activation.cpp

@@ -278,7 +278,7 @@ struct activation_impl : public typed_primitive_impl<activation> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/broadcast.cpp

@@ -112,7 +112,7 @@ struct broadcast_impl : public typed_primitive_impl<broadcast> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/concat.cpp

@@ -99,7 +99,7 @@ struct concatenation_impl : public typed_primitive_impl<concatenation> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/cpu_impl_helpers.hpp

@@ -77,8 +77,8 @@ using vector3D = vector2D<vector1D<T>>;
 template <typename T>
 using vector4D = vector2D<vector2D<T>>;
 
-inline event::ptr make_output_event(cldnn::stream& stream) {
-    if (stream.get_sync_method() == SyncMethods::events)
+inline event::ptr make_output_event(cldnn::stream& stream, bool is_output) {
+    if (is_output)
         return stream.create_user_event(true);
     return nullptr;
 }

src/plugins/intel_gpu/src/graph/impls/cpu/crop.cpp

@@ -99,7 +99,7 @@ struct crop_impl : public typed_primitive_impl<crop> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/detection_output.cpp

@@ -852,7 +852,7 @@ struct detection_output_impl : typed_primitive_impl<detection_output> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/eltwise.cpp

@@ -212,7 +212,7 @@ struct eltwise_impl : public typed_primitive_impl<eltwise> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/gather.cpp

@@ -102,7 +102,7 @@ struct gather_impl : public typed_primitive_impl<gather> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/non_max_suppression.cpp

@@ -416,7 +416,7 @@ struct non_max_suppression_impl : typed_primitive_impl<non_max_suppression> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     static std::unique_ptr<primitive_impl> create(const non_max_suppression_node&, const kernel_impl_params&) {
@@ -460,7 +460,7 @@ struct non_max_suppression_gather_impl : typed_primitive_impl<non_max_suppressio
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     static std::unique_ptr<primitive_impl> create(const non_max_suppression_gather_node&, const kernel_impl_params&) {

src/plugins/intel_gpu/src/graph/impls/cpu/proposal.cpp

@@ -435,7 +435,7 @@ struct proposal_impl : typed_primitive_impl<proposal> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache&, const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/range.cpp

@@ -79,7 +79,7 @@ struct range_impl : public typed_primitive_impl<range> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/read_value.cpp

@@ -81,7 +81,7 @@ struct read_value_impl : public typed_primitive_impl<read_value> {
             return stream.aggregate_events(res_events, res_events.size() > 1);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/reduce.cpp

@@ -139,7 +139,7 @@ struct reduce_impl : public typed_primitive_impl<reduce> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/reorder.cpp

@@ -72,7 +72,7 @@ struct reorder_impl : public typed_primitive_impl<reorder> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/scatter_update.cpp

@@ -94,7 +94,7 @@ struct scatter_update_impl : public typed_primitive_impl<scatter_update> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/select.cpp

@@ -89,7 +89,7 @@ struct select_impl : public typed_primitive_impl<select> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/shape_of.cpp

@@ -63,7 +63,7 @@ struct shape_of_impl : public typed_primitive_impl<shape_of> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/strided_slice.cpp

@@ -168,7 +168,7 @@ struct strided_slice_impl : public typed_primitive_impl<strided_slice> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/graph/impls/cpu/tile.cpp

@@ -96,7 +96,7 @@ struct tile_impl : public typed_primitive_impl<tile> {
             return stream.group_events(events);
         }
 
-        return make_output_event(stream);
+        return make_output_event(stream, instance.is_output());
     }
 
     void init_kernels(const kernels_cache& , const kernel_impl_params&) override {}

src/plugins/intel_gpu/src/plugin/graph.cpp

@@ -578,9 +578,11 @@ void Graph::update_profiling_info() {
         auto event = execIter->second;
         executedPrimitives.erase(execIter);
 
-        cldnn::instrumentation::profiling_info cldnnInfo{profiledID, event->get_profiling_info()};
+        if (event) {
+            cldnn::instrumentation::profiling_info cldnnInfo{profiledID, event->get_profiling_info()};
+            collectTimings(cldnnInfo, perfCount);
+        }
 
-        collectTimings(cldnnInfo, perfCount);
         perfCount.num++;
     }
 

src/plugins/intel_gpu/src/runtime/ocl/ocl_stream.cpp

@@ -392,6 +392,9 @@ void ocl_stream::wait_for_events(const std::vector<event::ptr>& events) {
     bool needs_barrier = false;
     std::vector<cl_event> clevents;
     for (auto& ev : events) {
+        if (!ev)
+            continue;
+
         if (auto ocl_base_ev = downcast<ocl_base_event>(ev.get())) {
             if (ocl_base_ev->get().get() != nullptr) {
                 clevents.push_back(ocl_base_ev->get().get());
@@ -411,9 +414,11 @@ void ocl_stream::wait_for_events(const std::vector<event::ptr>& events) {
         }
     }
 
-    auto err = clWaitForEvents(clevents.size(), &clevents[0]);
-    if (err != CL_SUCCESS) {
-        OPENVINO_THROW("[GPU] clWaitForEvents failed with ", err, " code");
+    if (!clevents.empty()) {
+        auto err = clWaitForEvents(static_cast<cl_uint>(clevents.size()), &clevents[0]);
+        if (err != CL_SUCCESS) {
+            OPENVINO_THROW("[GPU] clWaitForEvents failed with ", err, " code");
+        }
     }
 }
 

src/plugins/intel_gpu/src/runtime/stream.cpp

@@ -33,7 +33,7 @@ event::ptr stream::aggregate_events(const std::vector<event::ptr>& events, bool
     if (group && !is_output)
         return group_events(events);
 
-    return events.empty() ? nullptr : enqueue_marker(events, is_output);
+    return events.empty() ? (is_output ? create_user_event(true) : nullptr) : enqueue_marker(events, is_output);
 }
 
 }  // namespace cldnn

src/plugins/intel_gpu/tests/unit/module_tests/network_test.cpp

@@ -95,7 +95,7 @@ TEST(network_test, has_proper_event_for_in_order_queue) {
     net.set_input_data("input1", input_mem);
     net.execute();
 
-    ASSERT_FALSE(net.has_event("activation1"));
+    ASSERT_TRUE(net.has_event("activation1"));
     ASSERT_TRUE(net.has_event("concat"));
     ASSERT_TRUE(net.has_event("reorder"));
     ASSERT_TRUE(net.has_event("activation2"));

src/plugins/intel_gpu/tests/unit/module_tests/usm_memory_test.cpp

@@ -299,13 +299,11 @@ TEST_P(copy_between_gpu_buffer_and_gpu_usm, basic) {
         case allocation_type::usm_shared:
         case allocation_type::usm_device:
         {
-            auto ev = mem_dst->copy_from(stream, *usm_host_src, true);
-            ev->wait();
+            mem_dst->copy_from(stream, *usm_host_src, true);
             break;
         }
         case allocation_type::cl_mem: {
-            auto ev = mem_dst->copy_from(stream, *usm_host_src, true);
-            ev->wait();
+            mem_dst->copy_from(stream, *usm_host_src, true);
             break;
         }
         default: