performance: set timestamps as non-coherent since xe2

Related-To: NEO-14360

Signed-off-by: Szymon Morek <szymon.morek@intel.com>

shared/source/debug_settings/debug_variables_base.inl

@@ -383,11 +383,11 @@ DECLARE_DEBUG_VARIABLE(bool, DisableStatelessToStatefulOptimization, false, "Dis
 DECLARE_DEBUG_VARIABLE(bool, DisableConcurrentBlockExecution, false, "disables concurrent block kernel execution")
 DECLARE_DEBUG_VARIABLE(bool, UseNoRingFlushesKmdMode, true, "Windows only, passes flag to KMD that informs KMD to not emit any ring buffer flushes.")
 DECLARE_DEBUG_VARIABLE(bool, DisableZeroCopyForUseHostPtr, false, "When active all buffer allocations created with CL_MEM_USE_HOST_PTR flag will not share memory with CPU.")
-DECLARE_DEBUG_VARIABLE(bool, ForceNonCoherentModeForTimestamps, false, "When active timestamp buffers are allocated in non coherent memory.")
 DECLARE_DEBUG_VARIABLE(bool, SetAssumeNotInUse, true, "Set AssumeNotInUse flag in d3d destroy allocation.")
 DECLARE_DEBUG_VARIABLE(bool, MitigateHostVisibleSignal, false, "Reset host visible signal in CB events, flush L3 when synchronize")
 DECLARE_DEBUG_VARIABLE(bool, ForceZeroCopyForUseHostPtr, false, "When active all buffer allocations created with CL_MEM_USE_HOST_PTR flag will use share memory with CPU.")
 DECLARE_DEBUG_VARIABLE(bool, DummyPageBackingEnabled, false, "When true, pass page backing flag to KMD to recover from page faults. Windows only.");
+DECLARE_DEBUG_VARIABLE(int32_t, ForceNonCoherentModeForTimestamps, -1, "When active timestamp buffers are allocated in non coherent memory.")
 DECLARE_DEBUG_VARIABLE(int32_t, EnableReusingGpuTimestamps, -1, "Reuse GPU timestamp for next device time requests. -1: os-specific, 0: disable, 1: enable")
 DECLARE_DEBUG_VARIABLE(int32_t, AllowZeroCopyWithoutCoherency, -1, "Use cacheline flush instead of memory copy for map/unmap mem object")
 DECLARE_DEBUG_VARIABLE(int32_t, EnableHostPtrTracking, -1, "Enable host ptr tracking: -1 - default platform setting, 0 - disabled, 1 - enabled")

shared/source/gmm_helper/cache_settings_helper.cpp

@@ -99,13 +99,10 @@ GMM_RESOURCE_USAGE_TYPE_ENUM CacheSettingsHelper::getDefaultUsageTypeWithCaching
         return GMM_RESOURCE_USAGE_OCL_SYSTEM_MEMORY_BUFFER;
     case AllocationType::gpuTimestampDeviceBuffer:
     case AllocationType::timestampPacketTagBuffer:
-        if (debugManager.flags.ForceNonCoherentModeForTimestamps.get()) {
+        if (productHelper.isNonCoherentTimestampsModeEnabled()) {
             return GMM_RESOURCE_USAGE_OCL_BUFFER;
         }
-        if (productHelper.isDcFlushAllowed()) {
         return getDefaultUsageTypeWithCachingDisabled(allocationType, productHelper);
-        }
-        return GMM_RESOURCE_USAGE_OCL_BUFFER;
     default:
         return GMM_RESOURCE_USAGE_OCL_BUFFER;
     }
@@ -118,12 +115,6 @@ GMM_RESOURCE_USAGE_TYPE_ENUM CacheSettingsHelper::getDefaultUsageTypeWithCaching
     case AllocationType::internalHeap:
     case AllocationType::linearStream:
         return GMM_RESOURCE_USAGE_OCL_SYSTEM_MEMORY_BUFFER_CACHELINE_MISALIGNED;
-    case AllocationType::timestampPacketTagBuffer:
-    case AllocationType::gpuTimestampDeviceBuffer:
-        if (debugManager.flags.ForceNonCoherentModeForTimestamps.get()) {
-            return GMM_RESOURCE_USAGE_OCL_BUFFER;
-        }
-        [[fallthrough]];
     default:
         return productHelper.isNewCoherencyModelSupported() ? GMM_RESOURCE_USAGE_OCL_BUFFER_CSR_UC : GMM_RESOURCE_USAGE_OCL_BUFFER_CACHELINE_MISALIGNED;
     }

shared/source/os_interface/product_helper.h

@@ -269,6 +269,8 @@ class ProductHelper {
     virtual bool isCompressionForbidden(const HardwareInfo &hwInfo) const = 0;
     virtual bool isExposingSubdevicesAllowed() const = 0;
     virtual bool useAdditionalBlitProperties() const = 0;
+    virtual bool isNonCoherentTimestampsModeEnabled() const = 0;
+
     virtual bool getStorageInfoLocalOnlyFlag(LocalMemAllocationMode usmDeviceAllocationMode, bool defaultValue) const = 0;
     virtual ~ProductHelper() = default;
 

shared/source/os_interface/product_helper_before_xe2.inl

@@ -50,4 +50,12 @@ bool ProductHelperHw<gfxProduct>::isResourceUncachedForCS(AllocationType allocat
     return false;
 }
 
+template <PRODUCT_FAMILY gfxProduct>
+bool ProductHelperHw<gfxProduct>::isNonCoherentTimestampsModeEnabled() const {
+    if (debugManager.flags.ForceNonCoherentModeForTimestamps.get() != -1) {
+        return debugManager.flags.ForceNonCoherentModeForTimestamps.get();
+    }
+    return !this->isDcFlushAllowed();
+}
+
 } // namespace NEO

shared/source/os_interface/product_helper_hw.h

@@ -206,6 +206,7 @@ class ProductHelperHw : public ProductHelper {
     bool isCompressionForbidden(const HardwareInfo &hwInfo) const override;
     bool isExposingSubdevicesAllowed() const override;
     bool useAdditionalBlitProperties() const override;
+    bool isNonCoherentTimestampsModeEnabled() const override;
     bool getStorageInfoLocalOnlyFlag(LocalMemAllocationMode usmDeviceAllocationMode, bool defaultValue) const override;
 
     ~ProductHelperHw() override = default;

shared/source/os_interface/product_helper_xe2_and_later.inl

@@ -53,4 +53,12 @@ bool ProductHelperHw<gfxProduct>::isResourceUncachedForCS(AllocationType allocat
     return GraphicsAllocation::isAccessedFromCommandStreamer(allocationType);
 }
 
+template <PRODUCT_FAMILY gfxProduct>
+bool ProductHelperHw<gfxProduct>::isNonCoherentTimestampsModeEnabled() const {
+    if (debugManager.flags.ForceNonCoherentModeForTimestamps.get() != -1) {
+        return debugManager.flags.ForceNonCoherentModeForTimestamps.get();
+    }
+    return true;
+}
+
 } // namespace NEO

shared/test/common/mocks/mock_product_helper.cpp

@@ -483,6 +483,11 @@ bool ProductHelperHw<IGFX_UNKNOWN>::isResourceUncachedForCS(AllocationType alloc
     return false;
 }
 
+template <>
+bool ProductHelperHw<IGFX_UNKNOWN>::isNonCoherentTimestampsModeEnabled() const {
+    return false;
+}
+
 } // namespace NEO
 
 #include "shared/source/os_interface/product_helper.inl"

shared/test/common/test_files/igdrcl.config

@@ -621,7 +621,7 @@ ForceComputeWalkerPostSyncFlushWithWrite = -1
 DeferStateInitSubmissionToFirstRegularUsage = -1
 WaitForPagingFenceInController = -1
 DirectSubmissionPrintSemaphoreUsage = -1
-ForceNonCoherentModeForTimestamps = 0
+ForceNonCoherentModeForTimestamps = -1
 SetAssumeNotInUse = 1
 ExperimentalUSMAllocationReuseVersion = -1
 ForceNonWalkerSplitMemoryCopy = -1

shared/test/unit_test/gmm_helper/gmm_helper_tests.cpp

@@ -707,7 +707,7 @@ TEST(GmmTest, givenAllocationTypeWhenGettingUsageTypeThenReturnCorrectValue) {
                 break;
             case AllocationType::gpuTimestampDeviceBuffer:
             case AllocationType::timestampPacketTagBuffer:
-                expectedUsage = (forceUncached || productHelper.isDcFlushAllowed()) ? uncachedGmmUsageType
+                expectedUsage = (forceUncached || !productHelper.isNonCoherentTimestampsModeEnabled()) ? uncachedGmmUsageType
                                                                                                        : GMM_RESOURCE_USAGE_OCL_BUFFER;
                 break;
             case AllocationType::bufferHostMemory:
@@ -754,7 +754,7 @@ TEST(GmmTest, givenAllocationTypeWhenGettingUsageTypeThenReturnCorrectValue) {
                 break;
             case AllocationType::gpuTimestampDeviceBuffer:
             case AllocationType::timestampPacketTagBuffer:
-                expectedUsage = (forceUncached || productHelper.isDcFlushAllowed()) ? uncachedGmmUsageType
+                expectedUsage = (forceUncached || !productHelper.isNonCoherentTimestampsModeEnabled()) ? uncachedGmmUsageType
                                                                                                        : GMM_RESOURCE_USAGE_OCL_BUFFER;
                 break;
             case AllocationType::bufferHostMemory:
@@ -884,14 +884,25 @@ TEST(GmmTest, givenAllocationTypeAndMitigatedDcFlushWhenGettingUsageTypeThenRetu
     }
 }
 
-TEST(GmmTest, givenDebugFlagWhenTimestampAllocationsAreQueriedThenBufferPolicyIsReturned) {
+TEST(GmmTest, whenTimestampAllocationsAreQueriedThenCorrectBufferPolicyIsReturned) {
     DebugManagerStateRestore restorer;
-    debugManager.flags.ForceNonCoherentModeForTimestamps.set(1);
     MockExecutionEnvironment mockExecutionEnvironment{};
     const auto &productHelper = mockExecutionEnvironment.rootDeviceEnvironments[0]->getHelper<ProductHelper>();
-    auto expectedUsage = GMM_RESOURCE_USAGE_OCL_BUFFER;
+    auto uncachedType = productHelper.isNewCoherencyModelSupported() ? GMM_RESOURCE_USAGE_OCL_BUFFER_CSR_UC : GMM_RESOURCE_USAGE_OCL_BUFFER_CACHELINE_MISALIGNED;
+    auto expectedUsage = uncachedType;
+    if (productHelper.isNonCoherentTimestampsModeEnabled()) {
+        expectedUsage = GMM_RESOURCE_USAGE_OCL_BUFFER;
+    }
     EXPECT_EQ(expectedUsage, CacheSettingsHelper::getGmmUsageType(AllocationType::gpuTimestampDeviceBuffer, false, productHelper, defaultHwInfo.get()));
     EXPECT_EQ(expectedUsage, CacheSettingsHelper::getGmmUsageType(AllocationType::timestampPacketTagBuffer, false, productHelper, defaultHwInfo.get()));
+
+    debugManager.flags.ForceNonCoherentModeForTimestamps.set(0);
+    EXPECT_EQ(uncachedType, CacheSettingsHelper::getGmmUsageType(AllocationType::gpuTimestampDeviceBuffer, false, productHelper, defaultHwInfo.get()));
+    EXPECT_EQ(uncachedType, CacheSettingsHelper::getGmmUsageType(AllocationType::timestampPacketTagBuffer, false, productHelper, defaultHwInfo.get()));
+
+    debugManager.flags.ForceNonCoherentModeForTimestamps.set(1);
+    EXPECT_EQ(GMM_RESOURCE_USAGE_OCL_BUFFER, CacheSettingsHelper::getGmmUsageType(AllocationType::gpuTimestampDeviceBuffer, false, productHelper, defaultHwInfo.get()));
+    EXPECT_EQ(GMM_RESOURCE_USAGE_OCL_BUFFER, CacheSettingsHelper::getGmmUsageType(AllocationType::timestampPacketTagBuffer, false, productHelper, defaultHwInfo.get()));
 }
 
 TEST(GmmTest, givenForceAllResourcesUncachedFlagSetWhenGettingUsageTypeThenReturnUncached) {