Refactor addPipeControlWA, PipeControlWArequired

Related-To: NEO-3210
Change-Id: I0516154b323e29eeb697bf2253ca08ae1ce150d8
Signed-off-by: Krzysztof Gibala <[email protected]>

Author: kgibala

runtime/command_queue/gpgpu_walker.h

@@ -112,7 +112,8 @@ class GpgpuWalkerHelper {
 
     static void dispatchProfilingCommandsStart(
         TagNode<HwTimeStamps> &hwTimeStamps,
-        LinearStream *commandStream);
+        LinearStream *commandStream,
+        const HardwareInfo &hwInfo);
 
     static void dispatchProfilingCommandsEnd(
         TagNode<HwTimeStamps> &hwTimeStamps,
@@ -132,7 +133,8 @@ class GpgpuWalkerHelper {
         LinearStream *cmdStream,
         WALKER_TYPE<GfxFamily> *walkerCmd,
         TagNode<TimestampPacketStorage> *timestampPacketNode,
-        TimestampPacketStorage::WriteOperationType writeOperationType);
+        TimestampPacketStorage::WriteOperationType writeOperationType,
+        const HardwareInfo &hwInfo);
 
     static void dispatchScheduler(
         LinearStream &commandStream,

runtime/command_queue/gpgpu_walker_base.inl

@@ -104,7 +104,8 @@ void GpgpuWalkerHelper<GfxFamily>::addAluReadModifyWriteRegister(
 template <typename GfxFamily>
 void GpgpuWalkerHelper<GfxFamily>::dispatchProfilingCommandsStart(
     TagNode<HwTimeStamps> &hwTimeStamps,
-    LinearStream *commandStream) {
+    LinearStream *commandStream,
+    const HardwareInfo &hwInfo) {
 
     using MI_STORE_REGISTER_MEM = typename GfxFamily::MI_STORE_REGISTER_MEM;
 
@@ -113,7 +114,7 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchProfilingCommandsStart(
 
     PipeControlHelper<GfxFamily>::obtainPipeControlAndProgramPostSyncOperation(*commandStream,
                                                                                PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_TIMESTAMP,
-                                                                               timeStampAddress, 0llu, false);
+                                                                               timeStampAddress, 0llu, false, hwInfo);
 
     //MI_STORE_REGISTER_MEM for context local timestamp
     timeStampAddress = hwTimeStamps.getGpuAddress() + offsetof(HwTimeStamps, ContextStartTS);

runtime/command_queue/gpgpu_walker_bdw_plus.inl

@@ -171,19 +171,20 @@ void GpgpuWalkerHelper<GfxFamily>::setupTimestampPacket(
     LinearStream *cmdStream,
     WALKER_TYPE<GfxFamily> *walkerCmd,
     TagNode<TimestampPacketStorage> *timestampPacketNode,
-    TimestampPacketStorage::WriteOperationType writeOperationType) {
+    TimestampPacketStorage::WriteOperationType writeOperationType,
+    const HardwareInfo &hwInfo) {
 
     if (TimestampPacketStorage::WriteOperationType::AfterWalker == writeOperationType) {
         uint64_t address = timestampPacketNode->getGpuAddress() + offsetof(TimestampPacketStorage, packets[0].contextEnd);
         PipeControlHelper<GfxFamily>::obtainPipeControlAndProgramPostSyncOperation(*cmdStream,
-                                                                                   PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA, address, 0, false);
+                                                                                   PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA, address, 0, false, hwInfo);
     }
 }
 
 template <typename GfxFamily>
 size_t EnqueueOperation<GfxFamily>::getSizeRequiredCSKernel(bool reserveProfilingCmdsSpace, bool reservePerfCounters, CommandQueue &commandQueue, const Kernel *pKernel) {
     size_t size = sizeof(typename GfxFamily::GPGPU_WALKER) + HardwareCommandsHelper<GfxFamily>::getSizeRequiredCS(pKernel) +
-                  sizeof(PIPE_CONTROL) * (HardwareCommandsHelper<GfxFamily>::isPipeControlWArequired() ? 2 : 1);
+                  sizeof(PIPE_CONTROL) * (HardwareCommandsHelper<GfxFamily>::isPipeControlWArequired(pKernel->getDevice().getHardwareInfo()) ? 2 : 1);
     size += HardwareCommandsHelper<GfxFamily>::getSizeRequiredForCacheFlush(commandQueue, pKernel, 0U);
     size += PreemptionHelper::getPreemptionWaCsSize<GfxFamily>(commandQueue.getDevice());
     if (reserveProfilingCmdsSpace) {

runtime/command_queue/hardware_interface_base.inl

@@ -178,7 +178,7 @@ void HardwareInterface<GfxFamily>::dispatchKernelCommands(CommandQueue &commandQ
 
     if (commandQueue.getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
         auto timestampPacketNode = currentTimestampPacketNodes->peekNodes().at(currentDispatchIndex);
-        GpgpuWalkerHelper<GfxFamily>::setupTimestampPacket(&commandStream, nullptr, timestampPacketNode, TimestampPacketStorage::WriteOperationType::BeforeWalker);
+        GpgpuWalkerHelper<GfxFamily>::setupTimestampPacket(&commandStream, nullptr, timestampPacketNode, TimestampPacketStorage::WriteOperationType::BeforeWalker, commandQueue.getDevice().getHardwareInfo());
     }
 
     programWalker(commandStream, kernel, commandQueue, currentTimestampPacketNodes, dsh, ioh, ssh, globalWorkSizes,

runtime/command_queue/hardware_interface_bdw_plus.inl

@@ -57,7 +57,7 @@ inline void HardwareInterface<GfxFamily>::dispatchProfilingPerfStartCommands(
 
     // If hwTimeStampAlloc is passed (not nullptr), then we know that profiling is enabled
     if (hwTimeStamps != nullptr) {
-        GpgpuWalkerHelper<GfxFamily>::dispatchProfilingCommandsStart(*hwTimeStamps, commandStream);
+        GpgpuWalkerHelper<GfxFamily>::dispatchProfilingCommandsStart(*hwTimeStamps, commandStream, commandQueue.getDevice().getHardwareInfo());
     }
     if (hwPerfCounter != nullptr) {
         GpgpuWalkerHelper<GfxFamily>::dispatchPerfCountersCommandsStart(commandQueue, *hwPerfCounter, commandStream);
@@ -109,7 +109,7 @@ inline void HardwareInterface<GfxFamily>::programWalker(
 
     if (currentTimestampPacketNodes && commandQueue.getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
         auto timestampPacketNode = currentTimestampPacketNodes->peekNodes().at(currentDispatchIndex);
-        GpgpuWalkerHelper<GfxFamily>::setupTimestampPacket(&commandStream, walkerCmd, timestampPacketNode, TimestampPacketStorage::WriteOperationType::AfterWalker);
+        GpgpuWalkerHelper<GfxFamily>::setupTimestampPacket(&commandStream, walkerCmd, timestampPacketNode, TimestampPacketStorage::WriteOperationType::AfterWalker, commandQueue.getDevice().getHardwareInfo());
     }
 
     HardwareCommandsHelper<GfxFamily>::sendIndirectState(

runtime/command_stream/command_stream_receiver.h

@@ -93,6 +93,7 @@ class CommandStreamReceiver {
 
     LinearStream &getCS(size_t minRequiredSize = 1024u);
     OSInterface *getOSInterface() const { return osInterface; };
+    ExecutionEnvironment &getExecutionEnvironment() const { return executionEnvironment; };
 
     MOCKABLE_VIRTUAL void setTagAllocation(GraphicsAllocation *allocation);
     GraphicsAllocation *getTagAllocation() const {

runtime/command_stream/command_stream_receiver_hw_base.inl

@@ -179,7 +179,7 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
         auto address = getTagAllocation()->getGpuAddress();
         PipeControlHelper<GfxFamily>::obtainPipeControlAndProgramPostSyncOperation(commandStreamTask,
                                                                                    PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA,
-                                                                                   address, taskCount + 1, dispatchFlags.dcFlush);
+                                                                                   address, taskCount + 1, dispatchFlags.dcFlush, device.getHardwareInfo());
 
         this->latestSentTaskCount = taskCount + 1;
         DBG_LOG(LogTaskCounts, __FUNCTION__, "Line: ", __LINE__, "taskCount", taskCount);
@@ -496,7 +496,7 @@ inline void CommandStreamReceiverHw<GfxFamily>::flushBatchedSubmissions() {
 
         ResidencyContainer surfacesForSubmit;
         ResourcePackage resourcePackage;
-        auto pipeControlLocationSize = PipeControlHelper<GfxFamily>::getSizeForPipeControlWithPostSyncOperation();
+        auto pipeControlLocationSize = PipeControlHelper<GfxFamily>::getSizeForPipeControlWithPostSyncOperation(peekHwInfo());
         void *currentPipeControlForNooping = nullptr;
         void *epiloguePipeControlLocation = nullptr;
 

runtime/command_stream/experimental_command_buffer.inl

@@ -63,7 +63,7 @@ size_t ExperimentalCommandBuffer::getTimeStampPipeControlSize() noexcept {
     using PIPE_CONTROL = typename GfxFamily::PIPE_CONTROL;
 
     // Two P_C for timestamps
-    return 2 * PipeControlHelper<GfxFamily>::getSizeForPipeControlWithPostSyncOperation();
+    return 2 * PipeControlHelper<GfxFamily>::getSizeForPipeControlWithPostSyncOperation(*commandStreamReceiver->getExecutionEnvironment().getHardwareInfo());
 }
 
 template <typename GfxFamily>
@@ -73,7 +73,7 @@ void ExperimentalCommandBuffer::addTimeStampPipeControl() {
     uint64_t timeStampAddress = timestamps->getGpuAddress() + timestampsOffset;
 
     PipeControlHelper<GfxFamily>::obtainPipeControlAndProgramPostSyncOperation(*currentStream,
-                                                                               PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_TIMESTAMP, timeStampAddress, 0llu, false);
+                                                                               PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_TIMESTAMP, timeStampAddress, 0llu, false, *commandStreamReceiver->getExecutionEnvironment().getHardwareInfo());
 
     //moving to next chunk
     timestampsOffset += sizeof(uint64_t);

runtime/device_queue/device_queue_hw_base.inl

@@ -125,13 +125,13 @@ void DeviceQueueHw<GfxFamily>::addExecutionModelCleanUpSection(Kernel *parentKer
 
     PipeControlHelper<GfxFamily>::obtainPipeControlAndProgramPostSyncOperation(slbCS,
                                                                                PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA,
-                                                                               criticalSectionAddress, ExecutionModelCriticalSection::Free, false);
+                                                                               criticalSectionAddress, ExecutionModelCriticalSection::Free, false, device->getHardwareInfo());
 
     uint64_t tagAddress = reinterpret_cast<uint64_t>(device->getDefaultEngine().commandStreamReceiver->getTagAddress());
 
     PipeControlHelper<GfxFamily>::obtainPipeControlAndProgramPostSyncOperation(slbCS,
                                                                                PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA,
-                                                                               tagAddress, taskCount, false);
+                                                                               tagAddress, taskCount, false, device->getHardwareInfo());
 
     addMediaStateClearCmds();
 

runtime/gen9/hardware_commands_helper_gen9.cpp

@@ -16,7 +16,7 @@
 namespace NEO {
 
 template <>
-bool HardwareCommandsHelper<SKLFamily>::isPipeControlWArequired() { return true; }
+bool HardwareCommandsHelper<SKLFamily>::isPipeControlWArequired(const HardwareInfo &hwInfo) { return true; }
 
 template struct HardwareCommandsHelper<SKLFamily>;
 } // namespace NEO