Remove redundant value CsrSizeRequestFlags::numGrfRequiredChanged

Related-To: NEO-5995

Signed-off-by: Filip Hazubski <[email protected]>

Author: fhazubski-Intel

level_zero/core/source/cmdlist/cmdlist_hw_immediate.inl

@@ -33,7 +33,7 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::executeCommandListImm
         nullptr,                                                                  //flushStampReference
         NEO::QueueThrottle::MEDIUM,                                               //throttle
         this->getCommandListPreemptionMode(),                                     //preemptionMode
-        this->commandContainer.lastSentNumGrfRequired,                            //numGrfRequired
+        GrfConfig::NotApplicable,                                                 //numGrfRequired
         NEO::L3CachingSettings::l3CacheOn,                                        //l3CacheSettings
         this->requiredStreamState.stateComputeMode.threadArbitrationPolicy.value, //threadArbitrationPolicy
         NEO::AdditionalKernelExecInfo::NotApplicable,                             //additionalKernelExecInfo

opencl/test/unit_test/gen11/coherency_tests_gen11.cpp

@@ -46,19 +46,19 @@ struct Gen11CoherencyRequirements : public ::testing::Test {
 GEN11TEST_F(Gen11CoherencyRequirements, GivenSettingsWhenCoherencyRequestedThenProgrammingIsCorrect) {
     auto lriSize = sizeof(MI_LOAD_REGISTER_IMM);
     overrideCoherencyRequest(false, false);
-    EXPECT_FALSE(csr->isComputeModeNeeded());
+    EXPECT_FALSE(csr->streamProperties.stateComputeMode.isDirty());
 
     overrideCoherencyRequest(false, true);
-    EXPECT_FALSE(csr->isComputeModeNeeded());
+    EXPECT_FALSE(csr->streamProperties.stateComputeMode.isDirty());
 
     overrideCoherencyRequest(true, true);
     auto retSize = csr->getCmdSizeForComputeMode();
-    EXPECT_TRUE(csr->isComputeModeNeeded());
+    EXPECT_TRUE(csr->streamProperties.stateComputeMode.isDirty());
     EXPECT_EQ(lriSize, retSize);
 
     overrideCoherencyRequest(true, false);
     retSize = csr->getCmdSizeForComputeMode();
-    EXPECT_TRUE(csr->isComputeModeNeeded());
+    EXPECT_TRUE(csr->streamProperties.stateComputeMode.isDirty());
     EXPECT_EQ(lriSize, retSize);
 }
 

opencl/test/unit_test/gen12lp/coherency_tests_gen12lp.inl

@@ -72,19 +72,19 @@ GEN12LPTEST_F(Gen12LpCoherencyRequirements, GivenNoSharedHandlesWhenGettingCmdSi
     }
 
     overrideCoherencyRequest(false, false, false);
-    EXPECT_FALSE(csr->isComputeModeNeeded());
+    EXPECT_FALSE(csr->streamProperties.stateComputeMode.isDirty());
 
     overrideCoherencyRequest(false, true, false);
-    EXPECT_FALSE(csr->isComputeModeNeeded());
+    EXPECT_FALSE(csr->streamProperties.stateComputeMode.isDirty());
 
     overrideCoherencyRequest(true, true, false);
     auto retSize = csr->getCmdSizeForComputeMode();
-    EXPECT_TRUE(csr->isComputeModeNeeded());
+    EXPECT_TRUE(csr->streamProperties.stateComputeMode.isDirty());
     EXPECT_EQ(cmdsSize, retSize);
 
     overrideCoherencyRequest(true, false, false);
     retSize = csr->getCmdSizeForComputeMode();
-    EXPECT_TRUE(csr->isComputeModeNeeded());
+    EXPECT_TRUE(csr->streamProperties.stateComputeMode.isDirty());
     EXPECT_EQ(cmdsSize, retSize);
 }
 
@@ -99,19 +99,19 @@ GEN12LPTEST_F(Gen12LpCoherencyRequirements, GivenSharedHandlesWhenGettingCmdSize
     }
 
     overrideCoherencyRequest(false, false, true);
-    EXPECT_FALSE(csr->isComputeModeNeeded());
+    EXPECT_FALSE(csr->streamProperties.stateComputeMode.isDirty());
 
     overrideCoherencyRequest(false, true, true);
-    EXPECT_FALSE(csr->isComputeModeNeeded());
+    EXPECT_FALSE(csr->streamProperties.stateComputeMode.isDirty());
 
     overrideCoherencyRequest(true, true, true);
     auto retSize = csr->getCmdSizeForComputeMode();
-    EXPECT_TRUE(csr->isComputeModeNeeded());
+    EXPECT_TRUE(csr->streamProperties.stateComputeMode.isDirty());
     EXPECT_EQ(cmdsSize, retSize);
 
     overrideCoherencyRequest(true, false, true);
     retSize = csr->getCmdSizeForComputeMode();
-    EXPECT_TRUE(csr->isComputeModeNeeded());
+    EXPECT_TRUE(csr->streamProperties.stateComputeMode.isDirty());
     EXPECT_EQ(cmdsSize, retSize);
 }
 

shared/source/command_container/cmdcontainer.cpp

@@ -154,7 +154,6 @@ void CommandContainer::reset() {
 
     iddBlock = nullptr;
     nextIddInBlock = this->getNumIddPerBlock();
-    lastSentNumGrfRequired = 0;
     lastPipelineSelectModeRequired = false;
     lastSentUseGlobalAtomics = false;
 }

shared/source/command_container/cmdcontainer.h

@@ -77,7 +77,6 @@ class CommandContainer : public NonCopyableOrMovableClass {
 
     uint32_t slmSize = std::numeric_limits<uint32_t>::max();
     uint32_t nextIddInBlock = 0;
-    uint32_t lastSentNumGrfRequired = 0;
     bool lastPipelineSelectModeRequired = false;
     bool lastSentUseGlobalAtomics = false;
 

shared/source/command_stream/command_stream_receiver.h

@@ -384,7 +384,6 @@ class CommandStreamReceiver {
     std::chrono::microseconds gpuHangCheckPeriod{500'000};
     uint32_t lastSentL3Config = 0;
     uint32_t latestSentStatelessMocsConfig = 0;
-    uint32_t lastSentNumGrfRequired = GrfConfig::DefaultGrfNumber;
     uint64_t lastSentSliceCount = QueueSliceCount::defaultSliceCount;
 
     uint32_t requiredScratchSize = 0;

shared/source/command_stream/command_stream_receiver_hw.h

@@ -75,7 +75,6 @@ class CommandStreamReceiverHw : public CommandStreamReceiver {
     size_t getCmdSizeForComputeMode();
     MOCKABLE_VIRTUAL bool hasSharedHandles();
 
-    bool isComputeModeNeeded() const;
     bool isPipelineSelectAlreadyProgrammed() const;
     void programComputeMode(LinearStream &csr, DispatchFlags &dispatchFlags, const HardwareInfo &hwInfo);
 

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -268,10 +268,6 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
                                                                                                             dispatchFlags.pipelineSelectArgs.specialPipelineSelectMode,
                                                                                                             hwInfo);
 
-    if (dispatchFlags.numGrfRequired == GrfConfig::NotApplicable) {
-        dispatchFlags.numGrfRequired = lastSentNumGrfRequired;
-    }
-
     auto requiresCoherency = hwHelper.forceNonGpuCoherencyWA(dispatchFlags.requiresCoherency);
     this->streamProperties.stateComputeMode.setProperties(requiresCoherency, dispatchFlags.numGrfRequired,
                                                           dispatchFlags.threadArbitrationPolicy, hwInfo);
@@ -281,9 +277,6 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
     csrSizeRequestFlags.mediaSamplerConfigChanged = this->lastMediaSamplerConfig != static_cast<int8_t>(dispatchFlags.pipelineSelectArgs.mediaSamplerRequired);
     csrSizeRequestFlags.specialPipelineSelectModeChanged = isSpecialPipelineSelectModeChanged;
 
-    csrSizeRequestFlags.numGrfRequiredChanged = this->lastSentNumGrfRequired != dispatchFlags.numGrfRequired;
-    lastSentNumGrfRequired = dispatchFlags.numGrfRequired;
-
     csrSizeRequestFlags.activePartitionsChanged = isProgramActivePartitionConfigRequired();
 
     auto force32BitAllocations = getMemoryManager()->peekForce32BitAllocations();
@@ -809,7 +802,7 @@ size_t CommandStreamReceiverHw<GfxFamily>::getRequiredCmdStreamSize(const Dispat
     size += sizeof(typename GfxFamily::MI_BATCH_BUFFER_START);
 
     size += getCmdSizeForL3Config();
-    if (isComputeModeNeeded()) {
+    if (this->streamProperties.stateComputeMode.isDirty()) {
         size += getCmdSizeForComputeMode();
     }
     size += getCmdSizeForMediaSampler(dispatchFlags.pipelineSelectArgs.mediaSamplerRequired);
@@ -1373,7 +1366,7 @@ inline MemoryCompressionState CommandStreamReceiverHw<GfxFamily>::getMemoryCompr
 template <typename GfxFamily>
 inline bool CommandStreamReceiverHw<GfxFamily>::isPipelineSelectAlreadyProgrammed() const {
     const auto &hwInfoConfig = *HwInfoConfig::get(peekHwInfo().platform.eProductFamily);
-    return isComputeModeNeeded() && hwInfoConfig.is3DPipelineSelectWARequired() && isRcs();
+    return this->streamProperties.stateComputeMode.isDirty() && hwInfoConfig.is3DPipelineSelectWARequired() && isRcs();
 }
 
 template <typename GfxFamily>

shared/source/command_stream/command_stream_receiver_hw_tgllp_and_later.inl

@@ -9,25 +9,18 @@
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/device/device.h"
 #include "shared/source/helpers/pipe_control_args.h"
-#include "shared/source/helpers/state_compute_mode_helper.h"
 #include "shared/source/os_interface/hw_info_config.h"
 
 namespace NEO {
 template <typename GfxFamily>
 void CommandStreamReceiverHw<GfxFamily>::programComputeMode(LinearStream &stream, DispatchFlags &dispatchFlags, const HardwareInfo &hwInfo) {
-    if (isComputeModeNeeded()) {
+    if (this->streamProperties.stateComputeMode.isDirty()) {
         EncodeComputeMode<GfxFamily>::programComputeModeCommandWithSynchronization(
             stream, this->streamProperties.stateComputeMode, dispatchFlags.pipelineSelectArgs,
             hasSharedHandles(), hwInfo, isRcs());
     }
 }
 
-template <typename GfxFamily>
-inline bool CommandStreamReceiverHw<GfxFamily>::isComputeModeNeeded() const {
-    return StateComputeModeHelper<Family>::isStateComputeModeRequired(csrSizeRequestFlags, false) ||
-           this->streamProperties.stateComputeMode.isDirty();
-}
-
 template <>
 inline void CommandStreamReceiverHw<Family>::addPipeControlBeforeStateBaseAddress(LinearStream &commandStream) {
     PipeControlArgs args;

shared/source/command_stream/csr_definitions.h

@@ -123,7 +123,6 @@ struct CsrSizeRequestFlags {
     bool preemptionRequestChanged = false;
     bool mediaSamplerConfigChanged = false;
     bool hasSharedHandles = false;
-    bool numGrfRequiredChanged = false;
     bool specialPipelineSelectModeChanged = false;
     bool activePartitionsChanged = false;
 };