Add flag to control prefetcher disabling behaviour

Certain platforms might not require prefetcher to
be disabled in direct submission. This change
provides a way to control that behaviour.

Signed-off-by: Rafal Maziejuk <rafal.maziejuk@intel.com>
Related-To: NEO-7218

shared/source/debug_settings/debug_variables_base.inl

@@ -313,8 +313,9 @@ DECLARE_DEBUG_VARIABLE(int32_t, EnableRingSwitchTagUpdateWa, -1, "-1: default, 0
 DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionReadBackCommandBuffer, -1, "-1: default - disabled, 0 - disable, 1 - enable. If enabled, read first dword of cmd buffer after handling residency.")
 DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionReadBackRingBuffer, -1, "-1: default - disabled, 0 - disable, 1 - enable. If enabled, read first dword of ring buffer after handling residency.")
 DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionInsertExtraMiMemFenceCommands, -1, "-1: default, 0 - disable, 1 - enable. If enabled, add extra MI_MEM_FENCE instructions with acquire bit set")
-DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionInsertSfenceInstructionPriorToSubmission, -1, "-1: default, 0 - disable, 1 - Instert _mm_sfence before unlocking semaphore only, 2 - insert before and after semaphore")
+DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionInsertSfenceInstructionPriorToSubmission, -1, "-1: default, 0 - disable, 1 - Insert _mm_sfence before unlocking semaphore only, 2 - insert before and after semaphore")
 DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionMaxRingBuffers, -1, "-1: default, >0: max ring buffer count, During switch ring buffer, if there is no available ring, wait for completion instead of allocating new one if DirectSubmissionMaxRingBuffers is reached")
+DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionDisablePrefetcher, -1, "-1: default, 0 - disable, 1 - enable. If enabled, disable prefetcher is being dispatched")
 DECLARE_DEBUG_VARIABLE(bool, DirectSubmissionPrintBuffers, false, "Print address of submitted command buffers")
 
 /*FEATURE FLAGS*/

shared/source/direct_submission/direct_submission_hw.h

@@ -202,5 +202,6 @@ class DirectSubmissionHw {
     bool miMemFenceRequired = false;
     bool systemMemoryFenceAddressSet = false;
     bool completionFenceSupported = false;
+    bool isDisablePrefetcherRequired = false;
 };
 } // namespace NEO

shared/source/direct_submission/direct_submission_hw.inl

@@ -65,6 +65,11 @@ DirectSubmissionHw<GfxFamily, Dispatcher>::DirectSubmissionHw(const DirectSubmis
         disableCpuCacheFlush = disableCacheFlushKey == 1 ? true : false;
     }
 
+    isDisablePrefetcherRequired = hwInfoConfig->isPrefetcherDisablingInDirectSubmissionRequired();
+    if (DebugManager.flags.DirectSubmissionDisablePrefetcher.get() != -1) {
+        isDisablePrefetcherRequired = !!DebugManager.flags.DirectSubmissionDisablePrefetcher.get();
+    }
+
     UNRECOVERABLE_IF(!CpuInfo::getInstance().isFeatureSupported(CpuInfo::featureClflush) && !disableCpuCacheFlush);
 
     createDiagnostic();
@@ -298,7 +303,10 @@ template <typename GfxFamily, typename Dispatcher>
 inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeSemaphoreSection() {
     size_t semaphoreSize = EncodeSempahore<GfxFamily>::getSizeMiSemaphoreWait();
     semaphoreSize += getSizePrefetchMitigation();
-    semaphoreSize += 2 * getSizeDisablePrefetcher();
+
+    if (isDisablePrefetcherRequired) {
+        semaphoreSize += 2 * getSizeDisablePrefetcher();
+    }
 
     if (miMemFenceRequired) {
         semaphoreSize += MemorySynchronizationCommands<GfxFamily>::getSizeForSingleAdditionalSynchronizationForDirectSubmission(*hwInfo);

shared/source/direct_submission/direct_submission_prefetcher_pvc_and_later.inl

@@ -13,13 +13,15 @@ template <typename GfxFamily, typename Dispatcher>
 inline void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchDisablePrefetcher(bool disable) {
     using MI_ARB_CHECK = typename GfxFamily::MI_ARB_CHECK;
 
-    MI_ARB_CHECK arbCheck = GfxFamily::cmdInitArbCheck;
-    arbCheck.setPreParserDisable(disable);
+    if (isDisablePrefetcherRequired) {
+        MI_ARB_CHECK arbCheck = GfxFamily::cmdInitArbCheck;
+        arbCheck.setPreParserDisable(disable);
 
-    EncodeMiArbCheck<GfxFamily>::adjust(arbCheck);
+        EncodeMiArbCheck<GfxFamily>::adjust(arbCheck);
 
-    MI_ARB_CHECK *arbCheckSpace = ringCommandStream.getSpaceForCmd<MI_ARB_CHECK>();
-    *arbCheckSpace = arbCheck;
+        MI_ARB_CHECK *arbCheckSpace = ringCommandStream.getSpaceForCmd<MI_ARB_CHECK>();
+        *arbCheckSpace = arbCheck;
+    }
 }
 
-} // namespace NEO
+} // namespace NEO

shared/source/direct_submission/direct_submission_prefetcher_xe_hp_core_and_later.inl

@@ -13,13 +13,15 @@ template <typename GfxFamily, typename Dispatcher>
 inline void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchDisablePrefetcher(bool disable) {
     using MI_ARB_CHECK = typename GfxFamily::MI_ARB_CHECK;
 
-    MI_ARB_CHECK arbCheck = GfxFamily::cmdInitArbCheck;
-    arbCheck.setPreFetchDisable(disable);
+    if (isDisablePrefetcherRequired) {
+        MI_ARB_CHECK arbCheck = GfxFamily::cmdInitArbCheck;
+        arbCheck.setPreFetchDisable(disable);
 
-    EncodeMiArbCheck<GfxFamily>::adjust(arbCheck);
+        EncodeMiArbCheck<GfxFamily>::adjust(arbCheck);
 
-    MI_ARB_CHECK *arbCheckSpace = ringCommandStream.getSpaceForCmd<MI_ARB_CHECK>();
-    *arbCheckSpace = arbCheck;
+        MI_ARB_CHECK *arbCheckSpace = ringCommandStream.getSpaceForCmd<MI_ARB_CHECK>();
+        *arbCheckSpace = arbCheck;
+    }
 }
 
-} // namespace NEO
+} // namespace NEO

shared/source/os_interface/hw_info_config.h

@@ -133,6 +133,7 @@ class HwInfoConfig {
     virtual uint32_t getL1CachePolicy() const = 0;
     virtual bool isEvictionWhenNecessaryFlagSupported() const = 0;
     virtual void adjustNumberOfCcs(HardwareInfo &hwInfo) const = 0;
+    virtual bool isPrefetcherDisablingInDirectSubmissionRequired() const = 0;
 
     MOCKABLE_VIRTUAL ~HwInfoConfig() = default;
 
@@ -236,6 +237,7 @@ class HwInfoConfigHw : public HwInfoConfig {
     uint32_t getL1CachePolicy() const override;
     bool isEvictionWhenNecessaryFlagSupported() const override;
     void adjustNumberOfCcs(HardwareInfo &hwInfo) const override;
+    bool isPrefetcherDisablingInDirectSubmissionRequired() const override;
 
   protected:
     HwInfoConfigHw() = default;

shared/source/os_interface/hw_info_config.inl

@@ -482,4 +482,9 @@ uint32_t HwInfoConfigHw<gfxProduct>::getL1CachePolicy() const {
 
 template <PRODUCT_FAMILY gfxProduct>
 void HwInfoConfigHw<gfxProduct>::adjustNumberOfCcs(HardwareInfo &hwInfo) const {}
+
+template <PRODUCT_FAMILY gfxProduct>
+bool HwInfoConfigHw<gfxProduct>::isPrefetcherDisablingInDirectSubmissionRequired() const {
+    return true;
+}
 } // namespace NEO

shared/source/xe_hpc_core/os_agnostic_hw_info_config_xe_hpc_core.inl

@@ -53,3 +53,8 @@ void HwInfoConfigHw<gfxProduct>::getKernelExtendedProperties(uint32_t *fp16, uin
     *fp32 = (FP_ATOMIC_EXT_FLAG_GLOBAL_LOAD_STORE | FP_ATOMIC_EXT_FLAG_GLOBAL_ADD | FP_ATOMIC_EXT_FLAG_GLOBAL_MIN_MAX);
     *fp64 = (FP_ATOMIC_EXT_FLAG_GLOBAL_LOAD_STORE | FP_ATOMIC_EXT_FLAG_GLOBAL_ADD | FP_ATOMIC_EXT_FLAG_GLOBAL_MIN_MAX);
 }
+
+template <>
+bool HwInfoConfigHw<gfxProduct>::isPrefetcherDisablingInDirectSubmissionRequired() const {
+    return false;
+}

shared/test/common/mocks/mock_hw_info_config.cpp

@@ -406,6 +406,11 @@ uint32_t L1CachePolicyHelper<IGFX_UNKNOWN>::getDefaultL1CachePolicy() {
     return 0u;
 }
 
+template <>
+bool HwInfoConfigHw<IGFX_UNKNOWN>::isPrefetcherDisablingInDirectSubmissionRequired() const {
+    return true;
+}
+
 } // namespace NEO
 
 #include "shared/source/os_interface/hw_info_config.inl"

shared/test/common/test_files/igdrcl.config

@@ -446,4 +446,5 @@ ExperimentalEnableL0DebuggerForOpenCL = 0
 DebuggerDisableSingleAddressSbaTracking = 0
 ForceImagesSupport = -1
 ForceCsrLockInBcsEnqueueOnlyForGpgpuSubmission = -1
-ExperimentalEnableTileAttach = 0
+ExperimentalEnableTileAttach = 0
+DirectSubmissionDisablePrefetcher = -1

shared/test/unit_test/direct_submission/CMakeLists.txt

@@ -16,4 +16,8 @@ if(TESTS_XE_HP_CORE)
   target_sources(neo_shared_tests PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/direct_submission_tests_xe_hp_core.cpp)
 endif()
 
+if(TESTS_XE_HPC_CORE)
+  target_sources(neo_shared_tests PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/direct_submission_tests_xe_hpc_core.cpp)
+endif()
+
 add_subdirectories()

shared/test/unit_test/direct_submission/direct_submission_tests_1.cpp

@@ -926,7 +926,9 @@ HWTEST_F(DirectSubmissionTest,
     EXPECT_EQ(expectedStoreAddress, storeDataCmdAtPosition->getAddress());
 
     cmdBufferPosition += sizeof(MI_STORE_DATA_IMM);
-    cmdBufferPosition += directSubmission.getSizeDisablePrefetcher();
+    if (HwInfoConfig::get(defaultHwInfo->platform.eProductFamily)->isPrefetcherDisablingInDirectSubmissionRequired()) {
+        cmdBufferPosition += directSubmission.getSizeDisablePrefetcher();
+    }
     MI_SEMAPHORE_WAIT *semaphoreWaitCmdAtPosition = genCmdCast<MI_SEMAPHORE_WAIT *>(cmdBufferPosition);
     ASSERT_NE(nullptr, semaphoreWaitCmdAtPosition);
     EXPECT_EQ(COMPARE_OPERATION::COMPARE_OPERATION_SAD_GREATER_THAN_OR_EQUAL_SDD,
@@ -1103,3 +1105,24 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, DirectSubmissionTest, givenDebugFlagSetWhenDispatch
 
     EXPECT_EQ(0u, arbCheck->getPreParserDisable());
 }
+
+HWCMDTEST_F(IGFX_XE_HP_CORE, DirectSubmissionTest, givenDisablePrefetcherDebugFlagDisabledWhenDispatchingPrefetcherThenSetCorrectValue) {
+    DebugManagerStateRestore restore;
+    DebugManager.flags.DirectSubmissionDisablePrefetcher.set(0);
+
+    using MI_ARB_CHECK = typename FamilyType::MI_ARB_CHECK;
+    using Dispatcher = BlitterDispatcher<FamilyType>;
+
+    MockDirectSubmissionHw<FamilyType, Dispatcher> directSubmission(*pDevice->getDefaultEngine().commandStreamReceiver);
+
+    bool ret = directSubmission.allocateResources();
+    EXPECT_TRUE(ret);
+
+    directSubmission.dispatchDisablePrefetcher(true);
+
+    HardwareParse hwParse;
+    hwParse.parseCommands<FamilyType>(directSubmission.ringCommandStream, 0);
+    hwParse.findHardwareCommands<FamilyType>();
+    MI_ARB_CHECK *arbCheck = hwParse.getCommand<MI_ARB_CHECK>();
+    EXPECT_EQ(nullptr, arbCheck);
+}

shared/test/unit_test/direct_submission/direct_submission_tests_xe_hpc_core.cpp

@@ -0,0 +1,57 @@
+/*
+ * Copyright (C) 2022 Intel Corporation
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ */
+
+#include "shared/source/direct_submission/dispatchers/blitter_dispatcher.h"
+#include "shared/source/xe_hp_core/hw_cmds.h"
+#include "shared/test/common/cmd_parse/hw_parse.h"
+#include "shared/test/common/helpers/debug_manager_state_restore.h"
+#include "shared/test/common/mocks/mock_direct_submission_hw.h"
+#include "shared/test/common/test_macros/hw_test.h"
+#include "shared/test/unit_test/fixtures/direct_submission_fixture.h"
+
+using DirectSubmissionTestXeHpcCore = Test<DirectSubmissionFixture>;
+
+XE_HPC_CORETEST_F(DirectSubmissionTestXeHpcCore, givenXeHpcCoreWhenDispatchDisablePrefetcherIsCalledThenPrefetcherIsNotDisabled) {
+    using MI_ARB_CHECK = typename FamilyType::MI_ARB_CHECK;
+    using Dispatcher = BlitterDispatcher<FamilyType>;
+
+    MockDirectSubmissionHw<FamilyType, Dispatcher> directSubmission(*pDevice->getDefaultEngine().commandStreamReceiver);
+
+    EXPECT_EQ(sizeof(MI_ARB_CHECK), directSubmission.getSizeDisablePrefetcher());
+
+    bool ret = directSubmission.allocateResources();
+    EXPECT_TRUE(ret);
+
+    directSubmission.dispatchDisablePrefetcher(true);
+
+    HardwareParse hwParse;
+    hwParse.parseCommands<FamilyType>(directSubmission.ringCommandStream, 0);
+    hwParse.findHardwareCommands<FamilyType>();
+    MI_ARB_CHECK *arbCheck = hwParse.getCommand<MI_ARB_CHECK>();
+    EXPECT_EQ(nullptr, arbCheck);
+}
+
+XE_HPC_CORETEST_F(DirectSubmissionTestXeHpcCore, givenXeHpcCoreAndDisablePrefetcherDebugFlagEnabledWhenDispatchDisablePrefetcherIsCalledThenPrefetcherIsDisabled) {
+    DebugManagerStateRestore restore;
+    DebugManager.flags.DirectSubmissionDisablePrefetcher.set(1);
+
+    using MI_ARB_CHECK = typename FamilyType::MI_ARB_CHECK;
+    using Dispatcher = BlitterDispatcher<FamilyType>;
+
+    MockDirectSubmissionHw<FamilyType, Dispatcher> directSubmission(*pDevice->getDefaultEngine().commandStreamReceiver);
+
+    bool ret = directSubmission.allocateResources();
+    EXPECT_TRUE(ret);
+
+    directSubmission.dispatchDisablePrefetcher(true);
+
+    HardwareParse hwParse;
+    hwParse.parseCommands<FamilyType>(directSubmission.ringCommandStream, 0);
+    hwParse.findHardwareCommands<FamilyType>();
+    MI_ARB_CHECK *arbCheck = hwParse.getCommand<MI_ARB_CHECK>();
+    ASSERT_NE(nullptr, arbCheck);
+}

shared/test/unit_test/helpers/test_hw_info_config.cpp

@@ -161,3 +161,8 @@ HWTEST2_F(HwInfoConfigTest, givenPlatformWithUnsupportedL1CachePoliciesWhenGetL1
 
     EXPECT_EQ(0u, hwInfoConfig->getL1CachePolicy());
 }
+
+HWTEST_F(HwInfoConfigTest, givenHwInfoConfigWhenIsPrefetcherDisablingInDirectSubmissionRequiredThenTrueIsReturned) {
+    const auto &hwInfoConfig = *HwInfoConfig::get(defaultHwInfo->platform.eProductFamily);
+    EXPECT_TRUE(hwInfoConfig.isPrefetcherDisablingInDirectSubmissionRequired());
+}

shared/test/unit_test/xe_hpc_core/excludes_xe_hpc_core.cpp

@@ -37,3 +37,5 @@ HWTEST_EXCLUDE_PRODUCT(HwInfoConfigTest, whenCallingGetDeviceMemoryNameThenDdrIs
 HWTEST_EXCLUDE_PRODUCT(HwInfoConfigTest, givenHardwareInfoWhenCallingIsMaxThreadsForWorkgroupWARequiredThenFalseIsReturned, IGFX_XE_HPC_CORE);
 HWTEST_EXCLUDE_PRODUCT(HwInfoConfigTest, givenHwInfoConfigWhenAskedIfPipeControlPriorToNonPipelinedStateCommandsWARequiredThenFalseIsReturned, IGFX_XE_HPC_CORE);
 HWTEST_EXCLUDE_PRODUCT(HwInfoConfigTest, givenHwInfoConfigWhenAskedIfTile64With3DSurfaceOnBCSIsSupportedThenTrueIsReturned, IGFX_XE_HPC_CORE);
+HWTEST_EXCLUDE_PRODUCT(HwInfoConfigTest, givenHwInfoConfigWhenIsPrefetcherDisablingInDirectSubmissionRequiredThenTrueIsReturned, IGFX_XE_HPC_CORE);
+HWTEST_EXCLUDE_PRODUCT(DirectSubmissionTest, givenDebugFlagSetWhenDispatchingPrefetcherThenSetCorrectValue, IGFX_XE_HPC_CORE);

shared/test/unit_test/xe_hpc_core/pvc/test_hw_info_config_pvc.cpp

@@ -69,3 +69,8 @@ PVCTEST_F(PVCHwInfoConfig, givenHwInfoConfigWhenGettingEvictWhenNecessaryFlagSup
     const auto &hwInfoConfig = *HwInfoConfig::get(hwInfo.platform.eProductFamily);
     EXPECT_TRUE(hwInfoConfig.isEvictionWhenNecessaryFlagSupported());
 }
+
+PVCTEST_F(PVCHwInfoConfig, givenPVCHwInfoConfigWhenIsPrefetcherDisablingInDirectSubmissionRequiredThenFalseIsReturned) {
+    const auto &hwInfoConfig = *HwInfoConfig::get(productFamily);
+    EXPECT_FALSE(hwInfoConfig.isPrefetcherDisablingInDirectSubmissionRequired());
+}