Refactor CsrDependencies

Related-To: NEO-4175

Change-Id: If7e02235c1545e3e8ddf254c31523dd57286088d
Signed-off-by: Mateusz Hoppe <mateusz.hoppe@intel.com>

runtime/command_queue/enqueue_common.h

@@ -199,7 +199,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
     BlitPropertiesContainer blitPropertiesContainer;
 
     if (getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
-        csrDeps.fillFromEventsRequest(eventsRequest, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OnCsr);
+        eventsRequest.fillCsrDependencies(csrDeps, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OnCsr);
 
         size_t nodesCount = 0u;
         if (blitEnqueue || isCacheFlushCommand(commandType)) {
@@ -458,8 +458,8 @@ BlitProperties CommandQueueHw<GfxFamily>::processDispatchForBlitEnqueue(const Mu
     auto blitProperties = BlitProperties::constructProperties(blitDirection, *blitCommandStreamReceiver,
                                                               multiDispatchInfo.peekBuiltinOpParams());
     if (!queueBlocked) {
-        blitProperties.csrDependencies.fillFromEventsRequest(eventsRequest, *blitCommandStreamReceiver,
-                                                             CsrDependencies::DependenciesType::All);
+        eventsRequest.fillCsrDependencies(blitProperties.csrDependencies, *blitCommandStreamReceiver,
+                                          CsrDependencies::DependenciesType::All);
 
         blitProperties.csrDependencies.push_back(&timestampPacketDependencies.previousEnqueueNodes);
         blitProperties.csrDependencies.push_back(&timestampPacketDependencies.barrierNodes);
@@ -752,7 +752,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
     dispatchFlags.pipelineSelectArgs.specialPipelineSelectMode = specialPipelineSelectMode;
 
     if (getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
-        dispatchFlags.csrDependencies.fillFromEventsRequest(eventsRequest, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OutOfCsr);
+        eventsRequest.fillCsrDependencies(dispatchFlags.csrDependencies, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OutOfCsr);
         dispatchFlags.csrDependencies.makeResident(getGpgpuCommandStreamReceiver());
     }
 
@@ -944,7 +944,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueCommandWithoutKernel(
     );
 
     if (getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
-        dispatchFlags.csrDependencies.fillFromEventsRequest(eventsRequest, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OutOfCsr);
+        eventsRequest.fillCsrDependencies(dispatchFlags.csrDependencies, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OutOfCsr);
         dispatchFlags.csrDependencies.makeResident(getGpgpuCommandStreamReceiver());
     }
     CompletionStamp completionStamp = getGpgpuCommandStreamReceiver().flushTask(

runtime/command_stream/command_stream_receiver_hw.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2019 Intel Corporation
+ * Copyright (C) 2018-2020 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -11,7 +11,6 @@
 #include "core/helpers/options.h"
 #include "runtime/command_stream/command_stream_receiver.h"
 #include "runtime/execution_environment/execution_environment.h"
-#include "runtime/helpers/csr_deps.h"
 #include "runtime/helpers/dirty_state_helpers.h"
 
 namespace NEO {

runtime/command_stream/csr_definitions.h

@@ -6,13 +6,13 @@
  */
 
 #pragma once
+#include "core/command_stream/csr_deps.h"
 #include "core/command_stream/queue_throttle.h"
 #include "core/command_stream/thread_arbitration_policy.h"
 #include "core/helpers/hw_info.h"
 #include "core/helpers/pipeline_select_args.h"
 #include "core/kernel/grf_config.h"
 #include "core/memory_manager/memory_constants.h"
-#include "runtime/helpers/csr_deps.h"
 
 #include <limits>
 

runtime/helpers/CMakeLists.txt

@@ -16,8 +16,6 @@ set(RUNTIME_SRCS_HELPERS_BASE
   ${CMAKE_CURRENT_SOURCE_DIR}/built_ins_helper.h
   ${CMAKE_CURRENT_SOURCE_DIR}/cl_helper.h
   ${CMAKE_CURRENT_SOURCE_DIR}/convert_color.h
-  ${CMAKE_CURRENT_SOURCE_DIR}/csr_deps.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/csr_deps.h
   ${CMAKE_CURRENT_SOURCE_DIR}${BRANCH_DIR_SUFFIX}/device_helpers.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/dirty_state_helpers.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/dirty_state_helpers.h

runtime/helpers/blit_commands_helper.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2019 Intel Corporation
+ * Copyright (C) 2019-2020 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -197,8 +197,8 @@ void BlitProperties::setupDependenciesForAuxTranslation(BlitPropertiesContainer
 
     // wait for barrier and events before AuxToNonAux
     blitPropertiesContainer[0].csrDependencies.push_back(&timestampPacketDependencies.barrierNodes);
-    blitPropertiesContainer[0].csrDependencies.fillFromEventsRequest(eventsRequest, bcsCsr,
-                                                                     CsrDependencies::DependenciesType::All);
+    eventsRequest.fillCsrDependencies(blitPropertiesContainer[0].csrDependencies, bcsCsr,
+                                      CsrDependencies::DependenciesType::All);
 
     // wait for NDR before NonAuxToAux
     blitPropertiesContainer[numObjects].csrDependencies.push_back(&kernelTimestamps);

runtime/helpers/blit_commands_helper.h

@@ -1,14 +1,14 @@
 /*
- * Copyright (C) 2019 Intel Corporation
+ * Copyright (C) 2019-2020 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
  */
 
 #pragma once
+#include "core/command_stream/csr_deps.h"
 #include "core/memory_manager/memory_constants.h"
 #include "core/utilities/stackvec.h"
-#include "runtime/helpers/csr_deps.h"
 #include "runtime/helpers/properties_helper.h"
 
 #include <cstdint>

runtime/helpers/csr_deps.cpp

@@ -1,41 +1,15 @@
 /*
- * Copyright (C) 2019 Intel Corporation
+ * Copyright (C) 2019-2020 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
  */
 
-#include "runtime/helpers/csr_deps.h"
+#include "core/command_stream/csr_deps.h"
 
-#include "runtime/command_queue/command_queue.h"
-#include "runtime/event/event.h"
-#include "runtime/helpers/properties_helper.h"
 #include "runtime/helpers/timestamp_packet.h"
 
 namespace NEO {
-void CsrDependencies::fillFromEventsRequest(const EventsRequest &eventsRequest, CommandStreamReceiver &currentCsr,
-                                            DependenciesType depsType) {
-    for (cl_uint i = 0; i < eventsRequest.numEventsInWaitList; i++) {
-        auto event = castToObjectOrAbort<Event>(eventsRequest.eventWaitList[i]);
-        if (event->isUserEvent()) {
-            continue;
-        }
-
-        auto timestampPacketContainer = event->getTimestampPacketNodes();
-        if (!timestampPacketContainer || timestampPacketContainer->peekNodes().empty()) {
-            continue;
-        }
-
-        auto sameCsr = (&event->getCommandQueue()->getGpgpuCommandStreamReceiver() == &currentCsr);
-        bool pushDependency = (DependenciesType::OnCsr == depsType && sameCsr) ||
-                              (DependenciesType::OutOfCsr == depsType && !sameCsr) ||
-                              (DependenciesType::All == depsType);
-
-        if (pushDependency) {
-            this->push_back(timestampPacketContainer);
-        }
-    }
-}
 
 void CsrDependencies::makeResident(CommandStreamReceiver &commandStreamReceiver) const {
     for (auto &timestampPacketContainer : *this) {

runtime/helpers/csr_deps.h

@@ -1,30 +0,0 @@
-/*
- * Copyright (C) 2019 Intel Corporation
- *
- * SPDX-License-Identifier: MIT
- *
- */
-
-#pragma once
-#include "core/utilities/stackvec.h"
-
-namespace NEO {
-
-class TimestampPacketContainer;
-class CommandStreamReceiver;
-struct EventsRequest;
-
-class CsrDependencies : public StackVec<TimestampPacketContainer *, 32> {
-  public:
-    enum class DependenciesType {
-        OnCsr,
-        OutOfCsr,
-        All
-    };
-
-    void fillFromEventsRequest(const EventsRequest &eventsRequest, CommandStreamReceiver &currentCsr,
-                               DependenciesType depsType);
-
-    void makeResident(CommandStreamReceiver &commandStreamReceiver) const;
-};
-} // namespace NEO

runtime/helpers/hardware_commands_helper.h

@@ -17,7 +17,7 @@
 #include <cstdint>
 
 namespace NEO {
-
+class CommandQueue;
 class LinearStream;
 class IndirectHeap;
 struct CrossThreadInfo;

runtime/helpers/properties_helper.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2019 Intel Corporation
+ * Copyright (C) 2018-2020 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -7,12 +7,38 @@
 
 #include "runtime/helpers/properties_helper.h"
 
+#include "runtime/command_queue/command_queue.h"
 #include "runtime/helpers/mipmap.h"
+#include "runtime/helpers/timestamp_packet.h"
 #include "runtime/mem_obj/image.h"
 #include "runtime/mem_obj/mem_obj.h"
 #include "runtime/memory_manager/memory_manager.h"
 
 namespace NEO {
+
+void EventsRequest::fillCsrDependencies(CsrDependencies &csrDeps, CommandStreamReceiver &currentCsr, CsrDependencies::DependenciesType depsType) const {
+    for (cl_uint i = 0; i < this->numEventsInWaitList; i++) {
+        auto event = castToObjectOrAbort<Event>(this->eventWaitList[i]);
+        if (event->isUserEvent()) {
+            continue;
+        }
+
+        auto timestampPacketContainer = event->getTimestampPacketNodes();
+        if (!timestampPacketContainer || timestampPacketContainer->peekNodes().empty()) {
+            continue;
+        }
+
+        auto sameCsr = (&event->getCommandQueue()->getGpgpuCommandStreamReceiver() == &currentCsr);
+        bool pushDependency = (CsrDependencies::DependenciesType::OnCsr == depsType && sameCsr) ||
+                              (CsrDependencies::DependenciesType::OutOfCsr == depsType && !sameCsr) ||
+                              (CsrDependencies::DependenciesType::All == depsType);
+
+        if (pushDependency) {
+            csrDeps.push_back(timestampPacketContainer);
+        }
+    }
+}
+
 TransferProperties::TransferProperties(MemObj *memObj, cl_command_type cmdType, cl_map_flags mapFlags, bool blocking,
                                        size_t *offsetPtr, size_t *sizePtr, void *ptr, bool doTransferOnCpu)
     : memObj(memObj), ptr(ptr), cmdType(cmdType), mapFlags(mapFlags), blocking(blocking), doTransferOnCpu(doTransferOnCpu) {

runtime/helpers/properties_helper.h

@@ -7,6 +7,7 @@
 
 #pragma once
 
+#include "core/command_stream/csr_deps.h"
 #include "core/command_stream/queue_throttle.h"
 #include "runtime/api/cl_types.h"
 
@@ -21,7 +22,9 @@ struct EventsRequest {
     EventsRequest() = delete;
 
     EventsRequest(cl_uint numEventsInWaitList, const cl_event *eventWaitList, cl_event *outEvent)
-        : numEventsInWaitList(numEventsInWaitList), eventWaitList(eventWaitList), outEvent(outEvent){};
+        : numEventsInWaitList(numEventsInWaitList), eventWaitList(eventWaitList), outEvent(outEvent) {}
+
+    void fillCsrDependencies(CsrDependencies &csrDeps, CommandStreamReceiver &currentCsr, CsrDependencies::DependenciesType depsType) const;
 
     cl_uint numEventsInWaitList;
     const cl_event *eventWaitList;

runtime/helpers/task_information.cpp

@@ -7,6 +7,7 @@
 
 #include "runtime/helpers/task_information.h"
 
+#include "core/command_stream/csr_deps.h"
 #include "core/command_stream/linear_stream.h"
 #include "core/command_stream/preemption.h"
 #include "core/helpers/aligned_memory.h"
@@ -20,7 +21,6 @@
 #include "runtime/device/device.h"
 #include "runtime/device_queue/device_queue.h"
 #include "runtime/gtpin/gtpin_notify.h"
-#include "runtime/helpers/csr_deps.h"
 #include "runtime/helpers/enqueue_properties.h"
 #include "runtime/helpers/task_information.inl"
 #include "runtime/mem_obj/mem_obj.h"
@@ -228,7 +228,7 @@ CompletionStamp &CommandComputeKernel::submit(uint32_t taskLevel, bool terminate
     );
 
     if (timestampPacketDependencies) {
-        dispatchFlags.csrDependencies.fillFromEventsRequest(eventsRequest, commandStreamReceiver, CsrDependencies::DependenciesType::OutOfCsr);
+        eventsRequest.fillCsrDependencies(dispatchFlags.csrDependencies, commandStreamReceiver, CsrDependencies::DependenciesType::OutOfCsr);
         dispatchFlags.barrierTimestampPacketNodes = &timestampPacketDependencies->barrierNodes;
     }
     dispatchFlags.pipelineSelectArgs.specialPipelineSelectMode = kernel->requiresSpecialPipelineSelectMode();
@@ -275,7 +275,7 @@ void CommandWithoutKernel::dispatchBlitOperation() {
 
     UNRECOVERABLE_IF(kernelOperation->blitPropertiesContainer.size() != 1);
     auto &blitProperties = *kernelOperation->blitPropertiesContainer.begin();
-    blitProperties.csrDependencies.fillFromEventsRequest(eventsRequest, *bcsCsr, CsrDependencies::DependenciesType::All);
+    eventsRequest.fillCsrDependencies(blitProperties.csrDependencies, *bcsCsr, CsrDependencies::DependenciesType::All);
     blitProperties.csrDependencies.push_back(&timestampPacketDependencies->previousEnqueueNodes);
     blitProperties.csrDependencies.push_back(&timestampPacketDependencies->barrierNodes);
     blitProperties.outputTimestampPacket = currentTimestampPacketNodes->peekNodes()[0];
@@ -334,8 +334,7 @@ CompletionStamp &CommandWithoutKernel::submit(uint32_t taskLevel, bool terminate
 
     UNRECOVERABLE_IF(!commandStreamReceiver.peekTimestampPacketWriteEnabled());
 
-    dispatchFlags.csrDependencies.fillFromEventsRequest(eventsRequest, commandStreamReceiver, CsrDependencies::DependenciesType::OutOfCsr);
-
+    eventsRequest.fillCsrDependencies(dispatchFlags.csrDependencies, commandStreamReceiver, CsrDependencies::DependenciesType::OutOfCsr);
     makeTimestampPacketsResident(commandStreamReceiver);
 
     gtpinNotifyPreFlushTask(&commandQueue);

runtime/helpers/timestamp_packet.h

@@ -7,8 +7,8 @@
 
 #pragma once
 
+#include "core/command_stream/csr_deps.h"
 #include "core/helpers/non_copyable_or_moveable.h"
-#include "runtime/helpers/csr_deps.h"
 #include "runtime/helpers/hardware_commands_helper.h"
 #include "runtime/utilities/tag_allocator.h"