Add multiEngineQueue field to DispatchFlags and modify interfaces

Change-Id: Iaa4754a22e9b88201aed7df01c7d6e5fd06c84a9

Author: zzdanowicz

runtime/command_queue/enqueue_common.h

@@ -611,6 +611,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
     }
     dispatchFlags.numGrfRequired = numGrfRequired;
     dispatchFlags.specialPipelineSelectMode = specialPipelineSelectMode;
+    dispatchFlags.multiEngineQueue = this->multiEngineQueue;
     DEBUG_BREAK_IF(taskLevel >= Event::eventNotReady);
 
     if (gtpinIsGTPinInitialized()) {

runtime/command_stream/command_stream_receiver_hw.inl

@@ -326,7 +326,8 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
             newGSHbase,
             requiredL3Index,
             getMemoryManager()->getInternalHeapBaseAddress(),
-            device.getGmmHelper());
+            device.getGmmHelper(),
+            dispatchFlags);
 
         if (sshDirty) {
             StateBaseAddressHelper<GfxFamily>::programBindingTableBaseAddress(commandStreamCSR, ssh, stateBaseAddressCmdOffset,

runtime/command_stream/csr_definitions.h

@@ -31,6 +31,11 @@ constexpr int64_t maxTimeout = std::numeric_limits<int64_t>::max();
 }
 
 struct DispatchFlags {
+    CsrDependencies csrDependencies;
+    FlushStampTrackingObj *flushStampReference = nullptr;
+    QueueThrottle throttle = QueueThrottle::MEDIUM;
+    PreemptionMode preemptionMode = PreemptionMode::Disabled;
+    uint32_t numGrfRequired = GrfConfig::DefaultGrfNumber;
     bool blocking = false;
     bool dcFlush = false;
     bool useSLM = false;
@@ -39,14 +44,10 @@ struct DispatchFlags {
     bool mediaSamplerRequired = false;
     bool requiresCoherency = false;
     bool lowPriority = false;
-    QueueThrottle throttle = QueueThrottle::MEDIUM;
     bool implicitFlush = false;
     bool outOfOrderExecutionAllowed = false;
-    FlushStampTrackingObj *flushStampReference = nullptr;
-    PreemptionMode preemptionMode = PreemptionMode::Disabled;
-    CsrDependencies csrDependencies;
-    uint32_t numGrfRequired = GrfConfig::DefaultGrfNumber;
     bool specialPipelineSelectMode = false;
+    bool multiEngineQueue = false;
 };
 
 struct CsrSizeRequestFlags {

runtime/helpers/properties_helper.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018 Intel Corporation
+ * Copyright (C) 2018-2019 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -15,7 +15,7 @@ namespace OCLRT {
 class MemObj;
 class Buffer;
 
-enum class QueueThrottle {
+enum QueueThrottle : uint32_t {
     LOW,
     MEDIUM,
     HIGH

runtime/helpers/state_base_address.h

@@ -14,6 +14,7 @@ namespace OCLRT {
 class GmmHelper;
 class IndirectHeap;
 class LinearStream;
+struct DispatchFlags;
 
 template <typename GfxFamily>
 struct StateBaseAddressHelper {
@@ -27,7 +28,8 @@ struct StateBaseAddressHelper {
         uint64_t generalStateBase,
         uint32_t statelessMocsIndex,
         uint64_t internalHeapBase,
-        GmmHelper *gmmHelper);
+        GmmHelper *gmmHelper,
+        DispatchFlags &dispatchFlags);
 
     static void appendStateBaseAddressParameters(
         STATE_BASE_ADDRESS *stateBaseAddress,
@@ -36,7 +38,8 @@ struct StateBaseAddressHelper {
         const IndirectHeap &ssh,
         uint64_t generalStateBase,
         uint64_t internalHeapBase,
-        GmmHelper *gmmHelper);
+        GmmHelper *gmmHelper,
+        DispatchFlags &dispatchFlags);
 
     static void programBindingTableBaseAddress(LinearStream &commandStream, const IndirectHeap &ssh, size_t stateBaseAddressCmdOffset,
                                                GmmHelper *gmmHelper);

runtime/helpers/state_base_address.inl

@@ -22,7 +22,8 @@ void StateBaseAddressHelper<GfxFamily>::programStateBaseAddress(
     uint64_t generalStateBase,
     uint32_t statelessMocsIndex,
     uint64_t internalHeapBase,
-    GmmHelper *gmmHelper) {
+    GmmHelper *gmmHelper,
+    DispatchFlags &dispatchFlags) {
 
     auto pCmd = static_cast<STATE_BASE_ADDRESS *>(commandStream.getSpace(sizeof(STATE_BASE_ADDRESS)));
     *pCmd = GfxFamily::cmdInitStateBaseAddress;
@@ -57,7 +58,7 @@ void StateBaseAddressHelper<GfxFamily>::programStateBaseAddress(
     pCmd->setStatelessDataPortAccessMemoryObjectControlState(gmmHelper->getMOCS(statelessMocsIndex));
     pCmd->setInstructionMemoryObjectControlState(gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_STATE_HEAP_BUFFER));
 
-    appendStateBaseAddressParameters(pCmd, dsh, ioh, ssh, generalStateBase, internalHeapBase, gmmHelper);
+    appendStateBaseAddressParameters(pCmd, dsh, ioh, ssh, generalStateBase, internalHeapBase, gmmHelper, dispatchFlags);
 }
 
 template <typename GfxFamily>
@@ -68,7 +69,8 @@ void StateBaseAddressHelper<GfxFamily>::appendStateBaseAddressParameters(
     const IndirectHeap &ssh,
     uint64_t generalStateBase,
     uint64_t internalHeapBase,
-    GmmHelper *gmmHelper) {
+    GmmHelper *gmmHelper,
+    DispatchFlags &dispatchFlags) {
 }
 
 template <typename GfxFamily>

runtime/helpers/task_information.cpp

@@ -62,6 +62,7 @@ CompletionStamp &CommandMapUnmap::submit(uint32_t taskLevel, bool terminated) {
     dispatchFlags.lowPriority = cmdQ.getPriority() == QueuePriority::LOW;
     dispatchFlags.throttle = cmdQ.getThrottle();
     dispatchFlags.preemptionMode = PreemptionHelper::taskPreemptionMode(cmdQ.getDevice(), nullptr);
+    dispatchFlags.multiEngineQueue = cmdQ.isMultiEngineQueue();
 
     DEBUG_BREAK_IF(taskLevel >= Event::eventNotReady);
 
@@ -212,6 +213,8 @@ CompletionStamp &CommandComputeKernel::submit(uint32_t taskLevel, bool terminate
     dispatchFlags.throttle = commandQueue.getThrottle();
     dispatchFlags.preemptionMode = preemptionMode;
     dispatchFlags.mediaSamplerRequired = kernel->isVmeKernel();
+    dispatchFlags.multiEngineQueue = commandQueue.isMultiEngineQueue();
+
     if (commandStreamReceiver.peekTimestampPacketWriteEnabled()) {
         dispatchFlags.csrDependencies.fillFromEventsRequestAndMakeResident(eventsRequest, commandStreamReceiver, CsrDependencies::DependenciesType::OutOfCsr);
     }

runtime/mem_obj/buffer.h

@@ -163,6 +163,7 @@ class BufferHw : public Buffer {
                  zeroCopy, isHostPtrSVM, isObjectRedescribed) {}
 
     void setArgStateful(void *memory, bool forceNonAuxMode, bool disableL3Cache) override;
+    void appendBufferState(void *memory, Context *context, GraphicsAllocation *gfxAllocation);
 
     static Buffer *create(Context *context,
                           MemoryProperties properties,

runtime/mem_obj/buffer.inl

@@ -87,5 +87,11 @@ void BufferHw<GfxFamily>::setArgStateful(void *memory, bool forceNonAuxMode, boo
         surfaceState->setCoherencyType(RENDER_SURFACE_STATE::COHERENCY_TYPE_IA_COHERENT);
         surfaceState->setAuxiliarySurfaceMode(AUXILIARY_SURFACE_MODE::AUXILIARY_SURFACE_MODE_AUX_NONE);
     }
+
+    appendBufferState(memory, context, getGraphicsAllocation());
+}
+
+template <typename GfxFamily>
+void BufferHw<GfxFamily>::appendBufferState(void *memory, Context *context, GraphicsAllocation *gfxAllocation) {
 }
 } // namespace OCLRT

unit_tests/command_stream/command_stream_receiver_flush_task_2_tests.cpp

@@ -6,6 +6,7 @@
  */
 
 #include "reg_configs_common.h"
+#include "runtime/command_stream/csr_definitions.h"
 #include "runtime/helpers/hw_helper.h"
 #include "runtime/gmm_helper/gmm_helper.h"
 #include "runtime/helpers/state_base_address.h"
@@ -980,8 +981,17 @@ HWTEST_F(CommandStreamReceiverFlushTaskTests, givenCsrInNonDirtyStateAndBatching
 HWTEST_F(CommandStreamReceiverFlushTaskTests, givenCsrWhenGeneralStateBaseAddressIsProgrammedThenDecanonizedAddressIsWritten) {
     uint64_t generalStateBaseAddress = 0xffff800400010000ull;
     StateBaseAddressHelper<FamilyType> helper;
+    DispatchFlags dispatchFlags;
 
-    helper.programStateBaseAddress(commandStream, dsh, ioh, ssh, generalStateBaseAddress, 0, generalStateBaseAddress, pDevice->getGmmHelper());
+    helper.programStateBaseAddress(commandStream,
+                                   dsh,
+                                   ioh,
+                                   ssh,
+                                   generalStateBaseAddress,
+                                   0,
+                                   generalStateBaseAddress,
+                                   pDevice->getGmmHelper(),
+                                   dispatchFlags);
 
     HardwareParse hwParser;
     hwParser.parseCommands<FamilyType>(commandStream);